Genomic Selection For Crop Improvement - New Molecular Breeding Strategies For Crop Improvement (Springer, 2017)

Rajeev K.
Varshney
Manish Roorkiwal
Mark E. Sorrells Editors
Genomic
Selection for Crop
Improvement
New Molecular Breeding Strategies
for Crop Improvement
Genomic Selection for Crop Improvement
Rajeev K. Varshney • Manish Roorkiwal
Mark E. Sorrells
Editors
Genomic Selection for Crop

Improvement
New Molecular Breeding Strategies
for Crop Improvement
Editors
Rajeev K. Varshney Manish Roorkiwal
Center of Excellence in Genomics Center of Excellence in Genomics
Research Program - Genetic Gains Research Program - Genetic Gains
International Crops Research Institute International Crops Research Institute
for the Semi-Arid Tropics (ICRISAT) for the Semi-Arid Tropics (ICRISAT)
Patancheru, Telangana, India Patancheru, Telangana, India
Mark E. Sorrells
Department of Plant Breeding and Genetics
Cornell University
Ithaca, NY, USA
ISBN 978-3-319-63168-4 ISBN 978-3-319-63170-7 (eBook)

DOI 10.1007/978-3-319-63170-7
Library of Congress Control Number: 2017951639
© Springer International Publishing AG 2017

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission
or information storage and retrieval, electronic adaptation, computer software, or by similar or
dissimilar methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt
from the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, express or implied, with respect to the material contained
herein or for any errors or omissions that may have been made. The publisher remains neutral with
regard to jurisdictional claims in published maps and institutional affiliations.
Printed on acid-free paper
This Springer imprint is published by Springer Nature

The registered company is Springer International Publishing AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Foreword
Today the world is facing an unprecedented challenge: how to feed a growing

population predicted to reach over 9.1 billion people by 2050 on a resource base
threatened by climate change and with limited options for bringing new arable land
under cultivation. Associated challenges of high levels of women and child mal-
nutrition in Asia and sub-Saharan Africa and environmental degradation add to the
complexity threatening our future.
To meet these challenges, farmers need improved varieties of crops which give
higher productivity and economic returns while withstanding risks induced by
climate change such as high temperatures, changing spatial and temporal rainfall
distribution, and emerging pests and diseases. These new varieties must also
provide consumers, both rural and urban, with access to food that is highly
nutritious and safe.
A key task before the agricultural research community is to integrate genomics
into modern crop improvement to unlock the genetic diversity of food crops in ways
that maximize the availability of improved varieties with the range of production
and resilience traits (drought, heat, disease, and pest tolerance) alongside improved
nutritional value. Modern genomics provides new tools for increasing both the yield
and quality of crop products. Next-generation breeding will need to draw on
genomics as the “best bet” for sustainably eradicating hunger, malnutrition, and
poverty. Genomics coupled with advanced analytics and precision phenotyping can
dramatically increase our capacity to utilize genetic diversity and develop highly
nutritious, stress-tolerant crop varieties faster and cheaper than ever before and so
response with urgency to the realization of the Sustainable Development Goals by
2030.
Despite rates of genetic gain leveling off in many cropping systems, significant
efforts in genetic improvement have helped increase productivity and develop
climate-resilient varieties. Next-generation sequencing technologies are reducing
drastically the cost of genotyping and enabling genome-wide marker data to
support the design, development, and delivery of robust and nutritious crop
varieties.
v
vi Foreword
Genomic Selection for Crop Improvement is a timely resource to fill the gap
between genome science and crop breeding. In capturing the insights of global
leaders on genomics and crop improvement, I am confident that this resource will
advance our collective understanding and application of modern tools to unlock our
wealth of crop genetic diversity to deliver resilience and profitably for farmers and
nutritional value to consumers.
ICRISAT Dr. David Bergvinson

Patancheru, Telangana, India
October 17, 2017
Preface
The past decade has seen a tremendous shift toward using next-generation sequenc-
ing (NGS) technologies for development of powerful tools to identify underlying
genes for both simple and complex traits. The advent of NGS and high-throughput
genotyping technologies have reduced the genotyping cost significantly and made it
possible to use genome-wide marker data for prediction of phenotype to help reduce
the cost of phenotyping. Integration of genomics tools with conventional breeding
can forge new directions to meet environmental challenges efficiently in less time
and more accurately. First-generation molecular breeding approaches (marker-
assisted backcrossing (MABC) and marker-assisted recurrent selection (MARS))
require a lengthy process for developing genetic populations for identification of
linked markers for a few simply inherited traits but failed to improve complex traits
such as yield and drought tolerance due to their technical and genetic limitations. In
the case of complex traits which are generally controlled by large number of genes/
quantitative trait loci (QTLs) with small effect, “genomic selection (GS)” has
gained momentum in plant breeding due to the decline in the genotyping cost.
One of the strengths of GS lies in the ability to select an individual without
phenotypic data (predicting the individual’s breeding value) based on a prediction
model trained with phenotypes and genotypes. However, practicing GS is not as
simple as MABC and MARS and requires an understanding of complex statistical
models. GS has been widely used in cattle breeding and more recently has gained
popularity among plant breeders. This book is a timely effort to compile details
about GS for users providing basic as well as advanced understanding. The content
of this book will serve as a useful reference for users, covering the germplasm to be
used, phenotyping evaluation, marker genotyping methods, and statistical models
involved in genomic selection.
A total of 21 authors (Contributors) have contributed to the nine chapters of the
book. The editors of this volume are grateful to all the authors for their contribu-
tions and for their commendable effort in summarizing the published/unpublished
research work in a comprehensive, up-to-date manner. In addition, the cooperation
they have extended in terms of timely completion and revision of chapters from
vii
viii Preface
time to time is well appreciated. While editing this book, the strong support
received from many other colleagues (Drs. Aaron Lorenz, Isabel Vales, John M.
Hickey, and José Crossa) to review the chapters is greatly appreciated. Their
constructive comments and suggestions have been instrumental to further improve
the chapters.
The editors also would like to thank their respective families for their cooper-
ation and moral support as the editorial work for this book took away precious
moments that they should have spent together with their families. RKV is thankful
to Monika, his wife, for her constant encouragement and support and Prakhar (son)
and Preksha (daughter) for their love and cooperation. Similarly, MR is grateful to
his wife (Shweta) for her support and encouragement in doing editorial responsi-
bilities in addition to research duties at International Crops Research Institute for
the Semi-Arid Tropics (ICRISAT), with special thanks to Divit (son) for his
fondness. RKV and MR would also like to extend their sincerest thanks to
Dr. David J. Bergvinson, Director General, ICRISAT, and Dr. Peter S. Carberry,
Deputy Director General-Research, ICRISAT, for their help and support.
RKV and MR are also grateful to their colleagues from Center of Excellence in
Genomics (CEG), Research Program - Genetic Gains, ICRISAT, and the collabo-
rators for their direct/indirect suggestions during planning of the book. The coop-
eration and help received from Eric Stannard, Eric Hardy, and Rekha Udaiyar of
Springer during various stages of the development and completion of this book
project is gratefully acknowledged.
We hope that this book will be helpful and useful to students, young researchers,
and crop specialists.
Patancheru, Telangana, India Rajeev K. Varshney

Patancheru, Telangana, India Manish Roorkiwal
Ithaca, NY, USA Mark E. Sorrells
Contents
1 Genomic Selection for Crop Improvement: An Introduction . . . . . 1

Rajeev K. Varshney, Manish Roorkiwal, and Mark E. Sorrells
2 Training Population Design and Resource Allocation
for Genomic Selection in Plant Breeding . . . . . . . . . . . . . . . . . . . . 7
Aaron Lorenz and Liana Nice
3 Derivation of Linear Models for Quantitative Traits
by Bayesian Estimation with Gibbs Sampling . . . . . . . . . . . . . . . . . 23
Akihiro Nakaya and Sachiko Isobe
4 Bayesian Genomic-Enabled Prediction Models for Ordinal
and Count Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
Osval A. Montesinos-López, Abelardo Montesinos-López,
and José Crossa
5 Genomic Selection for Small Grain Improvement . . . . . . . . . . . . . 99
Jessica E. Rutkoski, Jared Crain, Jesse Poland,
and Mark E. Sorrells
6 Current Status and Prospects of Genomic Selection
in Legumes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
Ankit Jain, Manish Roorkiwal, Manish K. Pandey,
and Rajeev K. Varshney
7 Genomic Selection in Hybrid Breeding . . . . . . . . . . . . . . . . . . . . . . 149
Albert Wilhelm Schulthess, Yusheng Zhao,
and Jochen C. Reif
8 Opportunities and Challenges to Implementing Genomic
Selection in Clonally Propagated Crops . . . . . . . . . . . . . . . . . . . . . 185
Dorcus C. Gemenet and Awais Khan
ix
x Contents
9 Status and Perspectives of Genomic Selection in Forest

Tree Breeding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
Dario Grattapaglia
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251
Contributors
Jared Crain Department of Plant Pathology, Kansas State University, Manhattan,

KS, USA
José Crossa Biometric and Statistics Unit (BSU), International Maize and Wheat
Improvement Center (CIMMYT), México, D.F, Mexico
Dorcus C. Gemenet International Potato Centre (CIP), Lima, Peru
Dario Grattapaglia EMBRAPA Genetic Resources and Biotechnology – EPqB,
Brasilia, DF, Brazil
Universidade Católica de Brası́lia- SGAN, Brası́lia, DF, Brazil
Sachiko Isobe Laboratory of Plant Genomics and Genetics, Kazusa DNA
Research Institute, Chiba, Japan
Ankit Jain International Crops Research Institute for the Semi-Arid Tropics
(ICRISAT), Patancheru, Telangana, India
Awais Khan Plant Pathology and Plant-Microbe Biology Section, School of
Integrative Plant Science, Cornell University, New York State Agricultural Exper-
iment Station, Geneva, NY, USA
Aaron Lorenz University of Minnesota, Minneapolis, MN, USA
Abelardo Montesinos-López Departamento de Matemáticas, Centro
Universitario de Ciencias Exactas e Ingenierı́as (CUCEI), Universidad de
Guadalajara, Jalisco, Guadalajara, Mexico
Osval A. Montesinos-López Facultad de Telemática, Universidad de Colima,
Colima, Colima, Mexico
Akihiro Nakaya Department of Genome Informatics, Graduate School of Medi-
cine, Osaka University, Osaka, Japan
Liana Nice University of Minnesota, Minneapolis, MN, USA
xi
xii Contributors
Manish K. Pandey International Crops Research Institute for the Semi-Arid

Tropics (ICRISAT), Patancheru, Telangana, India
Jesse Poland Department of Plant Pathology, Kansas State University, Manhat-
tan, KS, USA
Jochen C. Reif Department of Breeding Research, Leibniz Institute of Plant
Genetics and Crop Plant Research (IPK), Gatersleben, Germany
Manish Roorkiwal International Crops Research Institute for the Semi-Arid
Jessica E. Rutkoski International Programs, College of Agriculture and Life
Sciences, Cornell University, Ithaca, NY, USA
Albert Wilhelm Schulthess Department of Breeding Research, Leibniz Institute
of Plant Genetics and Crop Plant Research (IPK), Gatersleben, Germany
Mark E. Sorrells Department of Plant Breeding and Genetics, Cornell University,
Ithaca, NY, USA
Rajeev K. Varshney International Crops Research Institute for the Semi-Arid
Yusheng Zhao Department of Breeding Research, Leibniz Institute of Plant
Genetics and Crop Plant Research (IPK), Gatersleben, Germany
Chapter 1
Genomic Selection for Crop Improvement:
An Introduction
Rajeev K. Varshney, Manish Roorkiwal, and Mark E. Sorrells
1.1 Introduction
Producing sufficient food to meet the demand of vastly growing population and
eradication of rural poverty is one of the critically important issues that the world is
facing. At the current pace, the world population is expected to cross the mark of
nine billion people by 2050 adding further pressure to already exhausted food
production systems. Considering the increasingly volatile climate, it will be diffi-
cult to maintain the crop production in conjugation with the demand, resulting in
increased food prices affecting people who already spend the highest percentage of
their disposable income on food. In addition to climate change, limited water
resource availability and poor soil health have the potential to restrict food crop
production. Furthermore, with increases in the world population, the availability of
agricultural land is decreasing. Under these constraints, to meet the rising demand
for food, agricultural production must increase by an estimated 50% without greatly
increasing water usage or expanding the total land area dedicated to agriculture.
Smallholder farmers, especially from underdeveloped and developing countries
with limited access to agricultural inputs or agricultural markets, are likely to be
affected by rising production costs and climate volatility. As per the United
R.K. Varshney (*) • M. Roorkiwal

International Crops Research Institute for the Semi-Arid Tropics (ICRISAT), Patancheru,
Telangana, India
e-mail: R.K.Varshney@CGIAR.ORG
M.E. Sorrells
Department of Plant Breeding and Genetics, Cornell University, Ithaca, New York 14853,
USA
© Springer International Publishing AG 2017 1

R.K. Varshney et al. (eds.), Genomic Selection for Crop Improvement,
DOI 10.1007/978-3-319-63170-7_1
2 R.K. Varshney et al.
Nations’ estimates, more than 790 million people globally do not have access to
sufficient nutritious food (https://sustainabledevelopment.un.org/sdg2), posing a
threat to achieving the Sustainable Development Goal (SDG) target of zero hunger
(universal access to safe, nutritious, and sufficient food at all times of the year).
In the event of these challenges, there is a need to look for new ways of breeding
for food crops and other plant species by using modern technologies. Modern
breeding approaches that have the capability to reduce breeding cycle time provide
more precision in selection, and more efficient use of genetic variation can be
exploited to increase the rate of genetic gains in breeding programs. The rapid
decline in the cost of sequencing and genotyping has led to the development of new
tools and strategies that can transform the way we breed plant species. In the past,
the cost of genotyping restricted the regular use of markers in breeding. In most
cases a limited number of markers for the target regions were used for selecting the
lines based on presence or absence of agriculturally important alleles. Development
of crop varieties using conventional breeding approaches has been effective but
time-consuming and labor-intensive. Recent advances in the next-generation
sequencing (NGS) technologies have been able to reduce the cost of genotyping
and sequencing. This has enabled the use of the high-throughput and cost-effective
high-density genotyping. These low-cost genotyping platforms have accelerated the
use of markers in the breeding programs using genome-wide approaches (Varshney
et al. 2014).
Integrating genomic tools with conventional breeding can have a major impact
for dealing with current and future environmental challenges more efficiently. In
such conditions, germplasm, genetic, and genomic resources are mandatory in all
plant species for rapid genetic gains in productivity of these species using decision
support tools. First-generation molecular breeding approaches (marker-assisted
backcrossing, marker-assisted recurrent selection) followed a lengthy process for
developing mapping populations for identification of markers linked to quantitative
trait loci (QTL) for a few simple traits. The majority of economically important
traits such as drought tolerance and yield are polygenic in nature and controlled by
multiple genes with small effects. In order to improve complex traits, such as
drought tolerance and yield, the modern breeding approach, genomic selection
(GS) (Meuwissen et al. 2001), can be deployed which specifically aims at improv-
ing quantitative traits by using genome-wide marker data without requiring iden-
tification of markers associated with QTL for traits of interest. GS uses a “training
population” of individuals that have been both phenotyped and genotyped to train a
prediction model for calculating genomic estimated breeding values (GEBVs).
Subsequently by using this model, GEBVs can be calculated for untested individ-
uals from a “candidate population”, and selection candidates (SCs) for making
crosses or for advanced yield trials can be identified. Although GEBVs do not
identify the function of the underlying QTLs/genes for the trait, they are an
excellent selection criterion (Jannink et al. 2010). GS attempts to capture the total
additive genetic variance with genome-wide marker coverage and effect estimates
(Rutkoski et al. 2011). Therefore, selection of an individual without phenotypic
data can be performed based on the individual’s predicted breeding value.
1 Genomic Selection for Crop Improvement: An Introduction 3
The models can be used to calculate GEBVs that help the breeder to identify
offspring that will be good parents in the next generation, based solely on genotypic
information about an existing line. The use of GEBVs in the context of genome-
wide prediction promises to help accelerate the rate of genetic gain in breeding.
The purpose of this book is to bring up-to-date information on GS breeding and
its application for crop species improvement. The editors believe that this book can
serve as ready reference for geneticists and crop breeders. This chapter introduces
the book and provides a summary of different chapters included in the book.
1.2 Methodologies and Models for GS
The first step toward deploying GS in crop breeding is to define a training set, which
should be closely related to the selection candidate population. Chapter 2 entitled
“Training population design and resource allocation for genomic selection in crop
breeding” provides detailed information about composition and optimization of
training population design related to population and trait architecture. In this
chapter, Aaron Lorenz and Liana Nice highlight the importance of the training
population design for predicting the breeding value of lines. The chapter focuses on
the process to select a calibration set (training population) for model training and
optimizes the resource allocation for field trials. With the advent of new technol-
ogies, it has become possible to collect phenotyping data in a more precise manner
with decreased error and increased efficiency and in larger quantities. NGS tech-
nologies are contributing to a continuing decrease in the genotyping cost and are
enabling the prediction of breeding value using genome-wide marker profiling. This
chapter also discusses the possible resource allocation in terms of the number of
replications for calibration of GS models vs allocation of more plots for model
training and allocation of plots within and across environment replication.
The Chap. 3 entitled “Derivation of linear models for quantitative traits by
Bayesian estimation with Gibbs sampling” contributed by Akihiro Nakaya and
Sachiko Isobe provides detailed information about construction of a prediction
model using a linear model. Model parameters are determined using the Bayesian
estimation with Gibbs sampling providing a theoretical background sufficient to
implement practical software for the model construction. The chapter also provides
a sample output by the implemented software. The chapter describes the different
prediction models including linear models, one-marker model, two-marker model
without interactions, and two-marker model with interactions to predict the trait
value of a sample using the environment types and genotypes. Prediction of the trait
values of samples based on their genetic and environmental factors is explained
using a prediction model that describes the relationship between the explanatory
factors observed in the samples and the trait values. This chapter suggests that
defining a prediction model for the target trait enables the selection to be based on
the predicted trait values, making it an essential part of genomic selection. When
the number of markers is greater than the number of samples, the prediction model
will be distorted. In order to address the issue related to model overfitting, detailed
inspection of the prediction model is necessary, and strategy based on the linear
mixed models and the Bayesian estimation may be useful in the prediction of trait
values of samples.
Montesinos-López and colleagues highlighted recent advances in models for
genomic-enabled prediction developed for ordinal categorical and count data in
Chap. 4 entitled “Bayesian genomic-enabled prediction models for ordinal and
count data”. Authors used these two models on simulated as well as a real dataset
using Bayesian framework suggesting that models used are a good alternative for
analyzing ordinal and count data in the context of genomic-enabled prediction.
Tested models have an advantage to perform an exact logistic or probit ordinal
regression without having to do approximations to perform a logistic ordinal
regression. Genotype (G) and environment (E) interaction is expected to affect
the prediction accuracies, and therefore modelling G E in the context of genomic-
enabled prediction plays a central role in crop breeding for the selection of
candidate genotypes. In order to best use GS models, understanding the data type
being analyzed is important before deciding on the modelling approach to be
employed.
1.3 GS in Field Crop Breeding
GS has been used or is being used in several crop breeding programs. This book
includes three chapters on applications offering both constraints and opportunities
of GS in crop breeding. The Chap. 5 entitled “Genomic selection for small grains
improvement” by Rutkoski and colleagues presents an overview of GS efforts being
undertaken in the small grain cereals. Authors in the chapter have explained
different approaches for implementation of GS in applied breeding programs.
A total of 40 GS studies have been undertaken so far in small grains including
wheat, barley, oat, rye, durum wheat, perennial ryegrass, and intermediate wheat-
grass. This chapter also discusses the factors affecting the GS prediction accuracies
in small grains and highlights the applicability of GS for analyzing and predicting
G E. They have discussed various scenarios affecting gain from selection and
cost relative to conventional breeding. Authors discussed the cost-benefit ratio for
deploying GS in cereal crops.
In Chap. 6 entitled “Current status and prospects of genomic selection in
legumes”, Jain and colleagues from ICRISAT provide an update on molecular
breeding in legumes and describe the ongoing GS efforts in some legume-breeding
programs including soybean, alfalfa, pea, chickpea, and groundnut. Legumes have
witnessed significant progress in the field of genomics and genetics in the past
decade, and efforts to deploy MAS have yielded some success for developing
superior legume varieties. However, as expected, MAS has not been that successful
for addressing complex traits such as drought and yield and therefore, efforts to
deploy GS in legume breeding were initiated. Authors have suggested that it is time
1 Genomic Selection for Crop Improvement: An Introduction 5
for other legumes to start deployment of GS in those breeding programs to achieve a

higher rate of genetic gain.
Hybrid breeding has been successful over varietal improvement in several crops.
Schulthess and colleagues from the Leibniz Institute of Plant Genetics and Crop
Plant Research (IPK), Germany, describe the basic concepts of hybrid breeding and
deployment of GS methods to simplify the philosophy underlying GS with hybrid
breeding in Chap. 7, “Genomic selection in hybrid breeding”. Authors have
explained the basic concepts relevant to hybrid breeding including dominance,
heterosis, combining abilities, and heterotic groups and patterns. The chapter also
describes the deployment of GS for hybrid genotypes using cross-validated predic-
tion accuracy, accommodating dominance effects within the GS model and other
GS approaches employed in hybrid breeding. Deployment of GS in hybrid breeding
is very challenging as many variables impact in hybrid breeding as compared to
pureline breeding. Authors propose an integrated plan with multidisciplinary skills
of breeders, scientists, and technicians before implementing GS in hybrid breeding.
1.4 GS for Improvement of Clonal Crops and Tree Species
Breeding in clonal crops and tree species is different from field crops. Therefore,
Gemenet and Khan in Chap. 8 entitled “Opportunities and challenges to
implementing genomic selection in clonally propagated crops” discuss issues
related to deployment of GS for improving the rate of genetic gain in clonal
crops. Authors highlight conventional breeding approaches for clonal crops that
involve crossing and planting of true seed plants in different generations followed
by evaluation of clones for several generations, making it a time- and resource-
consuming process. Therefore, GS-based selection of true seed plants can expedite
the breeding process. The chapter also describes the challenges including modelling
of genetic effects and heritability, linkage disequilibrium between markers and
QTLs, genetic architecture of traits, size of training population, and number of
generations following training model to deploy GS in clonal crops. For instance, GS
models generally handle additive effects and assume dominance and epistatic
effects as part of the residual which is not the case for clonally propagated crops,
as dominance and epistatic effects play an important role along with additive effects
and need special consideration. Therefore, for clonal crops, GS models with the
capability to include additive, dominance, and epistatic genetic effects need to be
employed for analysis.
For tree species, Dario Grattapaglia from EMBRAPA Genetic Resources and
Biotechnology, Brazil, provides perspectives of genomic selection and a compre-
hensive discussion on the factors relevant to GS in tree breeding in Chap. 9, “Status
and perspectives of genomic selection in forest tree breeding”. The chapter high-
lights the potential of GS in enhancing the rate of genetic gain in a tree breeding
program by reducing the selection cycle. In the case of a tree breeding program, the
long generation time typically necessary to complete a full breeding cycle can be
reduced by genotyping young seedlings and predicting their phenotype instead of

waiting for long a breeding cycle of 4–20 years or more. The authors have compiled
and presented all GS experimental studies in forest trees along with their key
attributes and performance of predictive abilities for different traits in the chapter.
1.5 Summary
As can be seen from the introduction of eight different chapters, GS, a modern
breeding approach, is gaining popularity and becoming the choice for many
breeders for improving complex traits. The book provides up-to-date information
about models, methodologies, factors affecting prediction accuracy, and some
examples of deployment of GS for crop improvement. This book will serve as
reference for users that provides basic as well as advanced understanding about
GS. The book is expected to serve as a useful review for users that explains the
germplasm to be used, phenotyping, marker genotyping methods, and statistical
models involved in GS. It also includes some examples of ongoing activities of
genomic selection in cereal, legume, clonal crop, and tree species.
References
Jannink JL, Lorenz AJ, Iwata H (2010) Genomic selection in plant breeding: from theory to
practice. Brief Funct Genomics 9:166–177
Meuwissen TH, Hayes BJ, Goddard ME (2001) Prediction of total genetic value using genome-
wide dense marker maps. Genetics 157:1819–1829
Rutkoski JE, Heffner EL, Sorrells ME (2011) Genomic selection for durable stem rust resistance in
wheat. Euphytica 179:161–173
Varshney RK, Terauchi R, McCouch SR (2014) Harvesting the promising fruits of genomics:
applying genome sequencing technologies to crop breeding. PLoS Biol 12(6):e1001883
Chapter 2
Training Population Design and Resource
Allocation for Genomic Selection in Plant
Breeding
Aaron Lorenz and Liana Nice
2.1 Introduction
Obtaining accurate and inexpensive estimates of genetic value is a fundamental

goal for plant breeders. To obtain these estimates and choose new varieties,
breeders continue to rely heavily on standard phenotyping practices for their
crops and traits of interest. Series of phenotypic testing procedures employed by
plant breeders can vary in scale, complexity, and relevance, both within and across
breeding programs. Scale can range from early generation, single plant observa-
tions to large, prerelease strip trials. Similarly, the complexity of phenotyping traits
within a breeding program can range from measuring flowering time, which can be
reliably phenotyped in a single environment in many cases, to drought tolerance
which can only be measured in a field setting if specific weather conditions occur or
if specially designed water stress nurseries are available. While phenotyping
followed by selection is the primary means of advancing lines, the time, cost, and
environmental error associated with obtaining phenotypic values leave room for
improvement. Advancements in phenotyping technologies have resulted in
decreased error, fewer inefficiencies in the phenotyping process, or larger quantities
of phenotypic data (Araus and Cairns 2014). An alternative yet complementary
approach to reducing phenotyping expenditures involves implementing genomic
selection using high-throughput molecular markers in breeding programs (Cabrera-
Bosquet et al. 2012).
Initially, molecular markers were used in the context of marker-assisted selec-
tion (MAS). This approach typically requires identification of tightly linked or
causal markers through mapping or cloning of quantitative trait loci (QTL) using
mapping populations or discovery panels. These markers are then used in parallel
A. Lorenz (*) • L. Nice

University of Minnesota, Minneapolis, MN 55455, USA
e-mail: lore0149@umn.edu

DOI 10.1007/978-3-319-63170-7_2
8 A. Lorenz and L. Nice
with measured phenotypes to make selections in the breeding program (Johnson

2004). The development of genomic selection techniques has altered the relation-
ship between markers and phenotypic data in breeding programs by introducing a
new role for phenotyping. Instead of using phenotypic data for direct measurement
of the phenotypic or breeding value of lines, phenotypic data in the context of
genomic selection is used to estimate marker effects and develop marker-based
predictive models. This is accomplished by developing calibration sets, or training
populations, that have been both phenotyped and genotyped with dense, genome-
wide markers. From this calibration set, a statistical model using all marker
information simultaneously is applied to predict the breeding values of individuals
that had not been phenotyped, known as the target population or prediction set.
With accurate predictive models, breeders can minimize the number of individuals
that are phenotyped and continue selection in environments that are not conducive
to obtaining quality phenotypes, such as off-season nurseries. Both scenarios can
effectively reduce the cost and/or time necessary for achieving the desired
genetic gain.
As genotyping costs continue to decrease, genomic selection will play an
increasingly important role in plant breeding. Research surrounding the hypothet-
ical and empirical implementation of genomic selection is an active field of study,
and the resulting techniques are being adopted by breeders in many crop species.
This movement toward an increasingly data-rich breeding process leads to ques-
tions surrounding the application of statistics, experimental design, and quantitative
genetics, to the selection of progenies for advancement and varietal release. While
the implementation of genomic selection may not affect the methods used for
phenotyping per se, breeders will need to consider training accurate genomic
prediction models when designing field trials, which would involve at least two
aspects: (1) selection of genotypes for field testing that are informative for model
building (i.e., training population design) in addition to those being advanced
toward variety release and (2) the allocation of field plots to genotypes. The
objective of this chapter is to review and discuss studies related to these two
important topics. It was our aim to provide the reader with a simple and hopefully
intuitive introduction to these topics.
2.2 Training Population Design
A critical first step toward the use of genomic selection is the establishment of the
training population (Jannink et al. 2010). Training population composition and the
way in which it’s established varies according to the role of genomic selection,
whether it be rapid recurrent selection within a closed population, selection within a
single biparental family, or selection among exotic plant accessions comprising a
germplasm collection. Approaches to compiling training populations include the
collection of new phenotypic data from targeted trials as well as the mining of
historical phenotypic data available on genotyped lines. Once again, the choice
2 Training Population Design and Resource Allocation for Genomic Selection. . . 9
between these two basic strategies depends upon the role of genomic selection in a
crop improvement program. The most important consideration of training popula-
tion design is the target population. In other words, the target population should be
defined first and foremost, and then the training population is designed around the
target population. There are two basic aims of training population design: (1) min-
imize costs associated with phenotyping by selecting smaller training populations,
and (2) maximize prediction accuracy for the set of individuals being predicted.
Balancing these goals should help breeders avoid poor prediction accuracies or
wasted resources.
To aid in this decision process, we review a range of studies that explore
composition and optimization of training population design. Windhausen et al.
(2012) laid out four breeding scenarios under which genomic selection may be
used: (1) training and target populations are segregating progenies from the same
cross, (2) training and target populations include related and unrelated genotypes,
(3) training and target sets include lines from a diverse germplasm collection, and
(4) recurrent selection within a closed synthetic population. Literature on training
population design for the first three scenarios will be reviewed. Literature on
training population design for the case of synthetic populations is sparse at the
present time; however one recently published study sheds light on this topic
(Schopp et al. 2017). Following the discussion of breeding scenarios, we explore
methods of training population selection and other considerations for training
population design related to population and trait architecture.
2.2.1 Training and Target Populations Are Segregating

Progenies from the Same Cross
The most straightforward way to conduct genomic selection is to create family-

specific training populations. In this scenario, individuals from the same family, or
biparental population, are used as both the training population and target popula-
tion. This approach has been discussed extensively in the maize breeding literature
(Bernardo and Yu 2007; Windhausen et al. 2012; Lorenz 2013; Jacobson et al.
2014), where large biparental families of inbred or doubled haploid lines are
common, as well as the wheat breeding literature (Heffner et al. 2011). To perform
genomic prediction, the entire family is genotyped, with a subset of these lines
serving as the training population to train a model to predict the individuals that
were not phenotyped. The genomic prediction model can also be used to predict
future selection cycles created by intermating selected individuals within the family
(Bernardo and Yu 2007; Combs and Bernardo 2013; Massman et al. 2013; Lorenz
2013). This breeding method is similar to marker-assisted recurrent selection in
terms of family structure (Johnson 2004), and the first published studies on genomic
selection for plant breeding used this approach (Whittaker et al. 2000; Bernardo and
Yu 2007).
Within-family predictions are often accurate, and only modest population sizes
and marker numbers are needed to achieve good prediction accuracy. High accu-
racy is possible because of the extensive linkage disequilibrium (LD) generated by
the initial hybridization event (Lorenzana and Bernardo 2009; Zhao et al. 2012).
This LD, which provides power for QTL mapping in biparental populations, also
leads to accurate predictions in the context of genomic selection. Generally, as
training population size increases within families, predictive ability increases
until a maximum has been reached. When working with high heritability traits,
the maximum prediction accuracy will be reached with a smaller training
population size.
In an era when genotyping can be less expensive than phenotyping, selecting a
subset of individuals to phenotype based on genotype data in order to reduce
population size (and thus cost of phenotyping) while maintaining QTL detection
power is a desirable goal. This is known as selective phenotyping. It has been
shown that selective phenotyping for QTL detection can enhance mapping power
and resolution depending on the number of QTL controlling a trait and their effect
sizes (Jannink 2005; Sen et al. 2009). Although the increase in power for QTL
mapping was minimal under optimized schemes, researchers have explored
whether similar optimizations could be used in genomic prediction. Marulanda
et al. (2015) simulated a biparental population with training population sets that
varied based on a large number of parameters. The parameters examined included
measures of collinearity among markers, LD, allele frequency, genetic relationships
among lines, diversity indices, mixed model parameters, and phenotypic variance
of the training population sets. While many of these factors varied with training
population size, none of the parameters derived from marker data were associated
with prediction accuracy. However, they did find that selection for enhanced
phenotypic variation of the training set led to greater prediction accuracy in the
case of smaller training populations. While marker-based optimization would be
ideal, the authors proposed that a first round of phenotyping with little replication
could be used for training population selection, followed by more intense
phenotyping of the optimized set across multiple locations (Marulanda et al.
2015). Ultimately, the lack of population structure in a biparental cross allows for
relatively good prediction from a random sample, as long as marker number and
population size are large enough to adequately train the selection model. The use of
genomic prediction to select non-phenotyped individuals within a single family,
however, needs to be carefully considered as studies on resource allocation have
suggested little to no benefit to only phenotyping a subset of a single family in order
to develop a model to predict the remaining individuals in a family, unless family
size is very large (Lorenz 2013; Endelman et al. 2014; Riedelsheimer and
Melchinger 2013).
2.2.2 Training and Target Populations Include Related

and Unrelated Genotypes
Realistically, models built from single biparental populations are limited in their
applications outside of breeding systems with easy access to large population sizes
and efficient doubled haploid technologies. The time required to develop and
phenotype biparental populations diminishes the potential time savings of
implementing genomic selection in place of phenotypic selection. Therefore,
methods that combine data across multiple related and/or unrelated families
would be valuable for breeders. This can be a challenge because many additional
factors come into play when combining data across populations, and adding more
individuals to the training population does not necessarily result in greater predic-
tion accuracy as we will discuss below.
The inclusion of related and unrelated genotypes in training and target
populations can be further broken down into two scenarios for our purposes here.
One scenario includes the development and testing of large families, often
consisting of DH lines, as is used in hybrid maize breeding. Families often consist
of 150 progenies or more. Under this scenario, it would be possible and appropriate
to pool together a few well-chosen families into a single training population. A
second common scenario is the development of many, small families. This scenario
is common in crops such as soybean and small grains, where crossing is followed by
multiple generations of inbreeding followed by visual selection on simply inherited
traits and on molecular markers tagging large-effect QTL. The number of progenies
per family reaching the yield trial phase is typically small (~20–40) which excludes
the possibility of within-family training populations as well as the pooling together
of only a few families to form a training population. Rather, training populations
would need to be formed by pooling together progenies that are derived from
various pedigrees and genetic backgrounds, spanning levels of relatedness. If the
populations have been genotyped, ancestral relationships among individuals in the
training and the target populations can be used to optimize the selection of
training set.
Numerous studies in both plant and animal breeding systems have shown that
prediction accuracy suffers when training populations are not related to the target
population (Pszczola et al. 2012; Windhausen et al. 2012; Ly et al. 2013; Technow
et al. 2013; Albrecht et al. 2014; Lorenz and Smith 2015). Analysis of genomic
selection in sheep showed that the strongest predictor of prediction accuracy of each
individual was the strength of relationship between the individual being predicted
and the top ten relatives in the training population (Clark et al. 2012). In contrast,
the mean relationship of the training population to the individual being predicted
was a weak predictor of prediction accuracy. Therefore, for an individual to be
predicted well using genomic prediction, the training population must include
several close relatives to that individual.
Along these same lines, results looking at pooling together large families (first
scenario described above) to predict a specific target family have generally
indicated that the best results are obtained when the families being pooled share one
parent with the target family. The addition of families sharing one parent with the
family-specific training population could increase model accuracy above the
family-specific training population, especially if the target family is small in size
(Schulz-Streeck et al. 2012; Jacobson et al. 2014). Lehermeier et al. (2014) found
that the predictive ability of pooled half-sib training populations could achieve
similar accuracy to family-specific training populations, but models built using
375 half-sib individuals were needed to reach the accuracy of models built using
only 50 full-sib individuals. Riedelsheimer et al. (2013) found that half-sib training
populations that shared one parent in common with the target population only
reached 50% of the predictive ability of family-specific training populations. This
study, however, only included a limited number of families (six), and in reality,
breeding programs would likely include many more families from which to
pool data.
The use of data from families unrelated to the target population (family) is more
problematic. Training populations consisting of only individuals unrelated to the
target population generally result in zero or near-zero prediction accuracy
(Riedelsheimer et al. 2013; Jacobson et al. 2014; Lehermeier et al. 2014). More-
over, the addition of unrelated families to a family-specific training population can
reduce prediction accuracy compared to the family-specific training population
alone (Riedelsheimer et al. 2013; Jacobson et al. 2014) or have no effect despite
increasing the training population size by up to sixfold (Zhao et al. 2012). Lorenz
and Smith (2015) showed a decline in prediction accuracy when individuals less
and less related to the target population were added to the training population.
Model accuracy was maximized by using smaller training populations that were
more closely related to the target population, and the addition of less related
individuals (mostly from a different breeding program) reduced accuracy of pre-
dictions for all traits. High marker densities may enhance the sharing of information
between families and improve prediction accuracy by pooling unrelated families
(Hickey et al. 2014). Hickey et al. (2014) found that training populations consisting
of families unrelated to the target family could produce models with accuracies
reaching 0.70, but only with population sizes approaching 20,000 individuals and
marker numbers greater than 10,000. It is possible that such training populations
could be constructed within the seed industry, but to our knowledge, nothing in the
public sector has yet come close to this scale.
2.2.3 Training and Target Populations Include Lines from

a Diverse Germplasm Collection
Besides predicting the genetic value of progenies comprising an active breeding

program, another role of genomic prediction includes the prediction of diverse
accessions comprising a germplasm collection. Germplasm collections can be very
large, containing up to hundreds of thousands of plant accessions. Advancements in

genotyping have made it possible to genotype entire germplasm collections (Hearne
et al. 2015; Song et al. 2015), opening up the possibility of predicting the perfor-
mance of all accessions (Jarquin et al. 2016). Phenotyping entire collections, on the
other hand, is often not feasible.
In this scenario, the training and target populations are essentially two subsets of
the same population, and thus the training population should be selected to repre-
sent the entire population. Several studies have examined the performance of
chosen statistical criteria and accompanying optimization algorithms in choosing
informative training populations.
Two criteria for assessing population design derived from mixed linear model
theory have been proposed: prediction error variance (PEV) and the generalized
coefficient of determination (CD). The PEV quantifies the error of prediction of
each random effect in the model. It is a function of the ratio of the model error to
genetic variance, the number of times an individual is measured, the number of
relatives of the individual included in the dataset, and the strength of their relation-
ship. The CD is defined as the amount of variation in true contrasts of genetic values
by predicted contrasts of genetic values, where the contrast is between each
individual being predicted in the target population and target population mean
(Laloë et al. 1996). Optimizing the reliability of these contrasts rather than of the
predictions per se takes the covariances among the individuals comprising the
target population into account and thus prevents the selection of closely related
individuals for training population formation (Rincent et al. 2012). Because genetic
variance is not included in the calculation of PEV, using this method may result in
selecting a relatively narrow training population that contains many close relatives.
These statistics are calculated for each individual in the target population, and the
average value across the target population (i.e., PEVmean and CDmean) is the final
optimization criteria.
Criteria related to minimizing PEV have been previously used to optimize data
collection in animal breeding programs (Laloë and Phocas 2003; Kuehn et al.
2007). Rincent et al. (2012) expanded the use of these criteria for training popula-
tion design and genomic selection in plant populations by implementing them in
combination with a simple exchange algorithm. An exchange algorithm involves
removal and replacement of one individual in the training population, followed by
calculation of the optimization criteria (e.g., PEVmean, CDmean) for the newly
formed training population. If the removal and replacement results in an improve-
ment measured by the chosen criteria, then the newly added individual remains;
else it is removed in place of another randomly sampled individual from the pool of
candidates. Rincent et al. (2012) found that an optimization scheme based on a
CDmean-optimized training population resulted in models of higher accuracy
compared to random sampling. An optimized population of 100 individuals
achieved the same prediction accuracy as a randomly selected population of
200, indicating large reduction in costs associated with phenotyping if this method
is applied. The CDmean criteria typically outperformed PEVmean and other diver-
sity criteria such as mean genetic relatedness of the selected training population
measured by the genomic relationship matrix. Isidro et al. (2015) applied these
same criteria to rice and wheat panels. These authors found that a simple, stratified
sampling method that ensured representation of each subpopulation in the training
set was superior for the highly structured rice population, whereas the CDmean
method was superior for the minimally structured wheat population. This indicates
that training population optimization does depend on the population, as well as the
trait.
Akdemir et al. (2015) also showed a consistent benefit to optimizing training
populations using relationship-based selection procedures. These authors focused
on a principal component-based approach that increased computational efficiency
and selected training populations with regard to a specified target population, rather
than relationships within the training population itself. Their results suggest that
such methods hold great potential to help choose maximally informative training
populations. Software that implement these methods have been made available to
the general user (Rincent et al. 2012; Akdemir et al. 2015).
2.2.4 Sources of Information and Population Genomic

Architecture Influence Training Population Design
The overall theme of the literature reviewed above is that relationships between
training and target populations are highly important for genomic prediction. It is
clear that small training populations can be used, and are likely superior, if they are
closely related to the target population. Very large training populations are needed
if little to no relationship exists. Some researchers (Campos de los et al. 2013;
Habier et al. 2013) have contributed a theoretical basis to the importance of
relationships and their interaction with marker density and prediction model. By
far the most common methods for performing genomic prediction are ridge regres-
sion best linear unbiased prediction (RR-BLUP) and genomic best linear unbiased
prediction (G-BLUP). Although these two models are mathematically equivalent
under the properties of the multivariate normal distribution (Habier et al. 2013),
practitioners of breeding and genomic selection view the information sharing of
these models from two different perspectives. From the RR-BLUP perspective,
information is shared between training populations and target populations through
the LD that exists between markers and QTL. Because of this, as marker-QTL LD
increases, prediction accuracy is expected to increase. From the G-BLUP perspec-
tive, information is shared via the realized genomic relationships of the training and
target individuals, which reflect the higher degree of resemblance of more closely
related individuals. Prediction of selection candidates is a function of the weighted
sum of phenotypes of individuals in the training population, with weights being
proportional to the genomic relationships (Campos de los et al. 2013). Depending
on the family structure and distribution of relationships, only a few close relatives
could be heavily weighted in the calculation of the genomic predictions, or weights
could be more uniformly distributed among individuals in training populations that

are distantly related to the target population.
Ultimately, it is the genetic relationships at causal loci that influence the
effectiveness of training populations to predict trait values in prediction sets and
not genetic relationships calculated according to markers (Habier et al. 2013;
Campos de los et al. 2013). The genomic relationship matrix, calculated using
genome-wide markers, is an estimate of the genomic relationship matrix at the
causal polymorphisms. Therefore, the accuracy of this estimation is what deter-
mines the effectiveness of G-BLUP (Campos de los et al. 2013). The resemblance
between the genomic relationship matrix at causal polymorphisms and the esti-
mated genomic relationship matrix based on markers is determined by marker-QTL
LD, which in turn is determined by pedigree relationships of the population,
population history and diversity, and marker density. Formula for calculating
PEV and reliability of predictions using expected genomic relationships based on
pedigree data was derived by Henderson (1975). Under these expectations, the
reliability of predictions approach 1.0 as the population size goes to infinity. This is
even the case if the training population is distantly related to the target population,
although the number of individuals required to increase accuracy is much higher
compared to the addition of more closely related individuals (Campos de los et al.
2013). Campos et al. (Campos de los et al. 2013) showed that the marker-QTL LD
sets an upper limit to prediction accuracy. This limit is lowered when there is a lack
of relationship between the training and target populations due to a decrease in
marker-QTL LD. This is especially true for distantly related individuals where
genomic relationships can be variable with respect to which markers are in high LD
with QTL, leading to a major source of error in the G-BLUP model (Hill and Weir
2011). The expected value of realized or pedigree relationship decreases, while the
variance of the realized relationship increases (Hill and Weir 2011).
Another way to look at this problem is by partitioning the information contained
in the genomic relationship matrix into three components: (1) marker-QTL LD,
which is an association between alleles among the population founders; (2) linkage
or co-segregation of alleles created by pedigree relationships at QTL; and (3) addi-
tive genetic relationships captured by markers (Habier et al. 2013). Habier et al.
(2013) used simulations and models to partition these three sources of information.
First, they showed that large population sizes and high marker densities are needed
to exploit the LD source of information. Secondly, the proportion of accuracy from
shared additive genetic relationships is reduced if training populations are expanded
by adding unrelated individuals. Accuracy due to LD, however, might be able to
compensate for low relatedness if very large training population sizes and/or high
marker densities are available. Still, Habier et al. (2013) present an example from
cattle data where the increase in the accuracy from LD could not compensate for the
loss of information from additive genetic relationships, and an overall decrease in
accuracy was observed after the addition of unrelated individuals. However, in their
maize example, additive genetic relationship accuracy was not changed by increas-
ing training population size, possibly due to a stronger family structure with many
more close relatives in the maize training population.
2.3 Resource Allocation for Phenotyping for Genomic

Prediction Model Calibration: To Rep or Not to Rep
A key design aspect of breeding programs is the allocation of resources among

breeding trials in terms of population size, number of replications, and locations.
Allocation decisions are multifaceted, involving consideration of trait logistics,
selection intensity, breeding stage of the materials being tested, and any associated
genotyping costs. These decisions affect the genetic gain that is possible, as well as
the power to detect QTL or accurately estimate marker effects. Considering selec-
tion in general, the fundamental trade-off is between achieving accurate estimates
of genotypic value by increasing replication and sampling a greater number of
individuals to increase the chance of identifying superior genotypes (Gauch and
Zobel 1996). Bos (1983) explored the optimum replication scheme for breeding
programs with respect to heritability. Because replication decreases phenotypic
variance, it also increases heritability. However, this increase only occurs to a point,
after which, fundamental changes to the experimental design would be needed to
improve heritability (Gauch and Zobel 1996). Therefore, more replication generally
results in better selection outcomes, with the exception of situations where herita-
bility is high and selection intensity is relaxed (Bos 1983). Gauch and Zobel (1996)
extended the scope of the Bos (1983) findings to consider the precision of data
collected and the relative efficiency of data collected. They found that in experi-
ments with high precision, adding replication beyond two is much less efficient than
in lower precision experiments that retain efficiency at greater replication numbers.
When considering markers, the focus changes from identifying the genotypic
value of individuals to estimating the additive genetic values of alleles. Knapp and
Bridges (1990) identified sources of variation in a QTL mapping experiment and
found that increasing population size instead of replication resulted in higher power
to detect QTL, particularly when residual genetic variation existed in the popula-
tion. Other studies reported similar findings, where larger population sizes gener-
ally result in higher power of QTL detection, and only moderately sized populations
of 150–300 individuals benefit from replication (Sch€on et al. 2004). Because of the
similarities between QTL mapping and MAS, resource allocation recommendations
for QTL mapping seem to transfer well to the context of MAS. Moreau et al.
(Moreau 2000) showed that larger population sizes resulted in maximum gain from
selection when traits were controlled by 5–10 QTL and when genotyping costs were
equal to phenotyping costs. The shift toward genomic selection has required a
reevaluation of these resource allocation recommendations in the context of a
cultivar development program.
In contrast to MAS, genomic selection aims to improve traits that are influenced
by many more QTL. In addition, because MAS considers marker effects as fixed
and statistical thresholds are used to determine which markers are used to calculate
marker scores, the success of MAS is closely related to QTL detection power. Here,
we will explore recent published literature that aims to address the resource
allocation questions relevant to genomic selection breeding programs and are not
sufficiently addressed by previous MAS studies. Specifically, we review: (1) the

value of replication for calibrating genomic selection models, (2) allocation of plots
to stages within the breeding cycle, and (3) allocation of plots to within versus
across environment replication.
2.3.1 Replication and Plot Allocation for Calibrating

Genomic Selection Models
To determine whether resource allocation recommendations for MAS can be

extrapolated to genomic selection models, Lorenz (2013) compared the accuracies
of genomic selection models (RR-BLUP) and MAS models (ordinary least squares,
OLS) under varying resource allocation schemes. The factors studied included total
plot budget, relative cost of genotyping in comparison to phenotyping, population
size, number of replications, heritability, and percentage of phenotyped individuals.
A very clear distinction in resource optimization between GS and MAS models was
found. Prediction accuracy was always substantially lower with MAS, and the
effect of replication was more apparent for MAS. Within a set budget, the addition
of replications and a consequent reduction of total individuals screened lead to a
decrease in accuracy with MAS. In contrast, the RR-BLUP model remained fairly
constant across different resource allocation scenarios, with low heritability, high
marker cost scenarios slightly favoring fewer individuals, and more replication.
When the total number of individuals was varied, the accuracy of genomic selection
models began to level off around 50–75 individuals, whereas MAS models took
many more individuals to achieve moderate prediction accuracies and continued to
improve as the numbers increased. These results suggest that the underlying
considerations for MAS are different from genomic selection.
2.3.2 Allocation of Resources Across Preliminary

and Advanced Breeding Tests
Breeding programs are generally structured with less replications in early genera-
tion screening, followed by greater replication, larger scale, and higher-cost trials in
later generations (Bernardo 2010). Breeders must take this tiered structure of the
breeding program into account when planning for genomic selection implementa-
tion. The stage at which genomic selection is implemented can affect genetic gain
as well as costs. Bassi et al. (2016) compared a series of wheat breeding schemes
that implemented genomic selection starting in generations F2, F3, F4, or F7. They
found that without including phenotypic selection at some stage in the program,
early generation F2 implementation had the highest potential for gain per year, but
also the highest genotyping costs. Longin et al. (2015) found that genomic selection
without a stage of phenotypic selection would only be useful with very high
prediction accuracies, possibly unrealistically high accuracies.
When accuracies are low, genomic selection can fill the role of a pretest,
whereby a low selection intensity is applied to remove the lowest performing
individuals (Longin et al. 2015). Most studies have focused on overall accuracy
of genomic selection, without considering the effectiveness of these selection
schemes to accurately remove the worst individuals or include the best. Endelman
et al. (2014) proposed using a response to selection metric Rmax based on the
maximum genotypic value of selections instead of Rmean based on mean values
for selection to analyze genetic gain in preliminary yield trials. Because the mean of
the selected population decreases as more individuals are selected, the Rmean
measure of genetic gain may encourage overly stringent selection in early gener-
ations that have less precise phenotypic estimates. Additional studies are needed to
expand on the use of genomic selection for early generation screening.
In contrast to early generation genomic selection, Bassi et al. (2016) compared
intermediate and later generation schemes. They found that implementation in the
F3 and F4 was a good compromise between no stage of phenotypic selection and the
minimal benefits of F7 implementation. While F7 implementation might be attrac-
tive to breeders because of its ease of implementation and lower genotyping costs,
this scheme resulted in minimal benefit over phenotypic selection alone. Longin
et al. (2015) concluded that for traits such as yield in wheat, with prediction
accuracies of approximately 0.3, one stage of genomic selection followed by one
stage of phenotypic selection provides the best compromise between genomic and
phenotypic selection.
2.3.3 Across Environment Versus Within Environment

Replication
For simplicity, much of the literature surrounding the topic of resource allocation
focuses on trade-offs within single environments, but the distribution of plot
resources across environments is a major consideration for breeders. Riedelsheimer
and Melchinger (2013) attempted to tackle this issue by developing a resource
allocation planning tool for distributing plot resources across and within environ-
ments. Their tool is limited to a single cycle of selection in biparental populations,
and it requires some degree of estimation based on previous experimental data.
Their calculations extend those developed by Daetwyler et al. (2008) to include
considerations of multi-environment testing. They found that larger budgets
favored more environments, with a lower proportion of plots being allocated to
the training set. As the budget decreased, the training set became a larger proportion
of the plots, and the number of environments tested decreased. Furthermore, they
emphasize that under low-budget scenarios, the optimization has much less
flexibility than under large-budget scenarios. Overall, their findings suggest rela-
tively few environments are needed for high prediction accuracy.
Endelman et al. (2014) looked at the effect of spreading replicates across
locations in preliminary yield trials under fixed budgets. They found that accuracy
increased as individuals were replicated across locations, but under a fixed budget,
the optimum accuracies were obtained without replication across locations, unless
the budget forced a relatively small training population size. That is, each individ-
ual should be phenotyped in only one environment, and population size should be
maximized to the extent the total number of plots across environments allows.
Markers provide the connectivity between environments. In contrast, across envi-
ronment phenotypic estimates based on phenotyping alone were poor when indi-
viduals were phenotyped in single environments. This reinforces the idea that
shared marker information does provide the connectivity between individuals,
providing potential cost savings for breeders implementing genomic selection.
2.3.4 Conclusions
The role of phenotyping in plant breeding is rapidly shifting from its previous sole
purpose of providing information for making breeding line advancement, parent
selection, and variety release decisions to providing the necessary data to train
genomic prediction models to enable genomic selection. As this new role of
phenotyping increases in relative importance, plant breeders need to rethink how
they design field trials, allocate plot resources to genotypes, and which individuals
are included in field trials. This review provides a short and simple introduction to
this literature. We have two basic conclusions at this time: (1) Training population
selection and design should take genetic relationships with the target population
into consideration, and optimization criteria such as PEVmean and CDmean com-
bined with exchange algorithms are useful methods for selecting training
populations. (2) The number of individuals phenotyped should be maximized by
allocating only one field plot to each genotype in most situations. Further research is
needed to develop a comprehensive theoretical framework for phenotyping for
genomic selection.
References
Akdemir D, Sanchez JI, Jannink J-L (2015) Optimization of genomic selection training
populations with a genetic algorithm. Genet Sel Evol 47:38. doi:10.1186/s12711-015-0116-6
Albrecht T, Auinger H-J, Wimmer V et al (2014) Genome-based prediction of maize hybrid
performance across genetic groups, testers, locations, and years. Theor Appl Genet
127:1375–1386. doi:10.1007/s00122-014-2305-z
Araus JL, Cairns JE (2014) Field high-throughput phenotyping: the new crop breeding frontier.
Trends Plant Sci 19:52–61. doi:10.1016/j.tplants.2013.09.008
Bassi FM, Bentley AR, Charmet G et al (2016) Breeding schemes for the implementation of
genomic selection in wheat (Triticum spp.) Plant Sci 242:23–36. doi:10.1016/j.plantsci.2015.
08.021
Bernardo R (2010) Breeding for quantitative traits in plants, 2nd edn. Stemma Press, Woodbury,
MN
Bernardo R, Yu J (2007) Prospects for Genomewide selection for quantitative traits in maize. Crop
Sci 47:1082–1090. doi:10.2135/cropsci2006.11.0690
Bos I (1983) The optimum number of replications when testing lines or families on a fixed number
of plots. Euphytica 32:311–318. doi:10.1007/BF00021439
Cabrera-Bosquet L, Crossa J, von Zitzewitz J et al (2012) High-throughput Phenotyping and
genomic selection: the Frontiers of crop breeding ConvergeF. J Integr Plant Biol 54:312–320.
doi:10.1111/j.1744-7909.2012.01116.x
Campos de los G, Vazquez AI, Fernando R et al (2013) Prediction of complex human traits using
the genomic best linear unbiased predictor. PLoS Genet 9:e1003608. doi:10.1371/journal.
pgen.1003608
Clark SA, Hickey JM, Daetwyler HD, van der Werf JH (2012) The importance of information on
relatives for the prediction of genomic breeding values and the implications for the makeup of
reference data sets in livestock breeding schemes. Genet Sel Evol 44:4. doi:10.1186/1297-
9686-44-4
Combs E, Bernardo R (2013) Accuracy of Genomewide selection for different traits with constant
population size, heritability, and number of markers. Plant Genome 6:0. doi: 10.3835/
plantgenome2012.11.0030
Daetwyler HD, Villanueva B, Woolliams JA (2008) Accuracy of predicting the genetic risk of
disease using a genome-wide approach. PLoS One 3:e3395. doi:10.1371/journal.pone.0003395
Endelman JB, Atlin GN, Beyene Y et al (2014) Optimal Design of Preliminary Yield Trials with
genome-wide markers. Crop Sci 54:48–59. doi:10.2135/cropsci2013.03.0154
Gauch HG, Zobel RW (1996) Optimal replication in selection experiments. Crop Sci 36:838–843.
doi:10.2135/cropsci1996.0011183X003600040002x
Habier D, Fernando RL, Garrick DJ (2013) Genomic BLUP decoded: a look into the black box of
genomic prediction. Genetics 194:597–607. doi:10.1534/genetics.113.152207
Hearne S, Franco J, Chen J et al (2015) Genome wide assessment of maize Genebank diversity;
synthesis of next generation technologies and GIS based approaches. San Diego, USA
Heffner EL, Jannink J-L, Iwata H, Souza E, Sorrells ME (2011) Genomic selection accuracy for
grain quality traits in biparental wheat populations. Crop Sci 51:2597–2606
Henderson CR (1975) Best linear unbiased estimation and prediction under a selection model.
Biometrics 31:423–447. doi:10.2307/2529430
Hickey JM, Dreisigacker S, Crossa J et al (2014) Evaluation of genomic selection training
population designs and genotyping strategies in plant breeding programs using simulation.
Crop Sci 54:1476–1488. doi:10.2135/cropsci2013.03.0195
Hill WG, Weir BS (2011) Variation in actual relationship as a consequence of Mendelian sampling
and linkage. Genet Res 93:47–64. doi:10.1017/S0016672310000480
Isidro J, Jannink J-L, Akdemir D et al (2015) Training set optimization under population structure
in genomic selection. Theor Appl Genet 128:145–158. doi:10.1007/s00122-014-2418-4
Jacobson A, Lian L, Zhong S, Bernardo R (2014) General combining ability model for
Genomewide selection in a Biparental cross. Crop Sci 54:895–905. doi:10.2135/cropsci2013.
11.0774
Jannink J-L (2005) Selective Phenotyping to accurately map quantitative trait loci. Crop Sci
45:901–908. doi:10.2135/cropsci2004.0278
Jannink J-L, Lorenz AJ, Iwata H (2010) Genomic selection in plant breeding: from theory to
practice. Brief Funct Genomics 9:166–177. doi:10.1093/bfgp/elq001
Jarquin D, Specht J, Lorenz A (2016) Prospects of genomic prediction in the USDA soybean
Germplasm collection: historical data creates robust models for enhancing selection of acces-
sions. G3 GenesGenomesGenetics:g3.116.031443. doi:10.1534/g3.116.031443
Johnson R (2004) Marker-assisted selection. Plant Breeding Reviews. John Wiley & Sons, In
Knapp SJ, Bridges WC (1990) Using molecular markers to estimate quantitative trait locus
parameters: power and genetic variances for Unreplicated and replicated progeny. Genetics
126:769–777
Kuehn LA, Notter DR, Nieuwhof GJ, Lewis RM (2007) Changes in connectedness over time in
alternative sheep sire referencing schemes. J Anim Sci 86:536–544. doi:10.2527/jas.2007-
0256
Laloë D, Phocas F (2003) A proposal of criteria of robustness analysis in genetic evaluation. Livest
Prod Sci 80:241–256. doi:10.1016/S0301-6226(02)00092-1
Laloë D, Phocas F, Ménissier F (1996) Considerations on measures of precision and connectedness
in mixed linear models of genetic evaluation. Genet Sel Evol 28:1–20. doi:10.1186/1297-9686-
28-4-359
Lehermeier C, Krämer N, Bauer E et al (2014) Usefulness of Multiparental populations of maize
(Zea mays L.) for genome-based prediction. Genetics 198:3–16. doi:10.1534/genetics.114.
161943
Longin CFH, Mi X, Würschum T (2015) Genomic selection in wheat: optimum allocation of test
resources and comparison of breeding strategies for line and hybrid breeding. Theor Appl
Genet 128:1297–1306. doi:10.1007/s00122-015-2505-1
Lorenz AJ (2013) Resource allocation for maximizing prediction accuracy and genetic gain of
genomic selection in plant breeding: a simulation experiment. G3 GenesGenomesGenetics
3:481–491. doi:10.1534/g3.112.004911
Lorenz AJ, Smith KP (2015) Adding genetically distant individuals to training populations reduces
genomic prediction accuracy in barley. Crop Sci 55:2657–2667. doi:10.2135/cropsci2014.12.
0827
Lorenzana RE, Bernardo R (2009) Accuracy of genotypic value predictions for marker-based
selection in biparental plant populations. Theor Appl Genet 120:151–161. doi:10.1007/
s00122-009-1166-3
Ly D, Hamblin M, Rabbi I et al (2013) Relatedness and genotype environment interaction affect
prediction accuracies in genomic selection: a study in cassava. Crop Sci 53:1312–1325. doi:10.
2135/cropsci2012.11.0653
Marulanda JJ, Melchinger AE, Würschum T (2015) Genomic selection in biparental populations:
assessment of parameters for optimum estimation set design. Plant Breed 134:623–630. doi:10.
1111/pbr.12317
Massman JM, Jung H-JG, Bernardo R (2013) Genomewide selection versus marker-assisted
recurrent selection to improve grain yield and Stover-quality traits for cellulosic ethanol in
maize. Crop Sci 53:58–66. doi:10.2135/cropsci2012.02.0112
Moreau L, Lemarie S, Charcosset A, Gallais A (2000) Economic efficiency of one cycle of
marker-assisted selection. Crop Sci 40:329–337. doi:10.2135/cropsci2000.402329x
Pszczola M, Strabel T, Mulder HA, Calus MPL (2012) Reliability of direct genomic values for
animals with different relationships within and to the reference population. J Dairy Sci
95:389–400. doi:10.3168/jds.2011-4338
Riedelsheimer C, Endelman JB, Stange M et al (2013) Genomic predictability of interconnected
Biparental maize populations. Genetics 194:493–503. doi:10.1534/genetics.113.150227
Riedelsheimer C, Melchinger AE (2013) Optimizing the allocation of resources for genomic
selection in one breeding cycle. Theor Appl Genet 126:2835–2848. doi:10.1007/s00122-013-
2175-9
Rincent R, Laloë D, Nicolas S et al (2012) Maximizing the reliability of genomic selection by
optimizing the calibration set of reference individuals: comparison of methods in two diverse
groups of maize Inbreds (Zea mays L.) Genetics 192:715–728. doi:10.1534/genetics.112.
141473
Sch€on CC, Utz HF, Groh S et al (2004) Quantitative trait locus mapping based on resampling in a
vast maize testcross experiment and its relevance to quantitative genetics for complex traits.
Genetics 167:485–498
Schopp P, Muller D, Technow F, Melchinger A (2017) Accuracy of genomic prediction in

synthetic populations depending on the number of parents, relatedness, and ancestral linkage
disequilibrium. Genetics 205:441–454
Schulz-Streeck T, Ogutu JO, Karaman Z et al (2012) Genomic selection using multiple
populations. Crop Sci 52:2453–2461. doi:10.2135/cropsci2012.03.0160
Sen Ś, Johannes F, Broman KW (2009) Selective genotyping and Phenotyping strategies in a
complex trait context. Genetics 181:1613–1626. doi:10.1534/genetics.108.094607
Song Q, Hyten DL, Jia G et al (2015) Fingerprinting soybean Germplasm and its utility in genomic
research. G3 GenesGenomesGenetics 5:1999–2006. doi:10.1534/g3.115.019000
Technow F, Bürger A, Melchinger AE (2013) Genomic prediction of northern corn leaf blight
resistance in maize with combined or separated training sets for Heterotic groups. G3
GenesGenomesGenetics 3:197–203. doi:10.1534/g3.112.004630
Whittaker JC, Thompson R, Denham MC (2000) Marker-assisted selection using ridge regression.
Genet Res 75:249–252. doi: null
Windhausen VS, Atlin GN, Hickey JM et al (2012) Effectiveness of genomic prediction of maize
hybrid performance in different breeding populations and environments. G3
GenesGenomesGenetics 2:1427–1436. doi:10.1534/g3.112.003699
Zhao Y, Gowda M, Liu W et al (2012) Accuracy of genomic selection in European maize elite
breeding populations. Theor Appl Genet 124:769–776. doi:10.1007/s00122-011-1745-y
Chapter 3
Derivation of Linear Models for Quantitative
Traits by Bayesian Estimation with Gibbs
Sampling
Akihiro Nakaya and Sachiko Isobe
3.1 Introduction
Prediction of the trait values of samples based on their genetic and environmental
factors is one of the vital steps in the genomic selection (GS) process used to
identify valuable individuals for the use in the breeding process (Meuwissen et al.
2001). The GS strategies are different from other marker-assisted selection (MAS)
strategies in that they use the predicted trait values. To obtain the estimated values
of a trait, a prediction model that describes the relationship between the observed
explanatory factors, e.g., the genetic and environmental factors, in the samples and
the trait values is derived using a training dataset consisting of the candidates of the
explanatory factors and the true values of the trait. Such a prediction model can be
composed of the effects by the explanatory factors selected, and when these effects
are thought to be additive, the trait value of a sample is estimated as their weighted
summation. The models based on the summation are referred to as linear models
and are suitable for focusing on the additive effects of the explanatory factors.
Aside from the genetic factors obtained by the genome-wide markers, the Bayesian
approach used in the derivation process is the key element for GS.
In this chapter, construction of a prediction model using a linear model and
determination of the model parameters using the Bayesian estimation with Gibbs
sampling are explained with an assuming dataset example for GS. Knowing the
A. Nakaya (*)
Department of Genome Informatics, Graduate School of Medicine, Osaka University,
Yamadaoka, 2-2, Suita, Osaka 565-0871, Japan
e-mail: nakaya@gi.med.osaka-u.ac.jp
S. Isobe
Laboratory of Plant Genomics and Genetics, Kazusa DNA Research Institute,
Kazusa-Kamatari, 2-6-7, Kisarazu, Chiba 292-0818, Japan
e-mail: sisobe@kazusa.or.jp

DOI 10.1007/978-3-319-63170-7_3
24 A. Nakaya and S. Isobe
details of the model derivation, one can understand what is carried out during the
prediction of the trait values. This will also provide a theoretical background
sufficient to implement practical software for the model construction. A sample
output by the implemented software is presented.
3.2 A Dataset for Genomic Selection
We assume N samples and M genetic markers in a given dataset. The ith sample
(i ¼ 1 , . . . , N ) is associated with a numerical trait value yi and the genotype of the
jth genetic marker zij ( j ¼ 1 , . . . , M ). A tuple of the values (yi, zi1, . . . , ziM) repre-
sents the data of the ith subject (i ¼ 1, . . . , N ). If the samples are also associated
with L values indicating whether they have the hth environment type xih
(h ¼ 1 , . . . , L ), then the tuple of the values for the ith subject is (yi, xi1, . . . , xiL,
zi1, . . . , ziM). The dataset can be represented by a vector of the trait values, the
profiles of the environment types in an N L matrix, and the profiles of the
genotypes in an N M matrix. We assume there is no missing data.
Figure 3.1 shows an example of the dataset. The dataset consists of three parts: P,
E, and G, respectively, corresponding to phenotypes (trait values), environment
types (L ¼ 2), and genotypes (M ¼ 50) of the samples (N ¼ 30). In this example, as
shown by the histogram in part P, the baseline values of yi follow N(0.75, 0.052) for
the upper 15 samples (i ¼ 1 , . . . , 15) and N(0.25, 0.052) for the lower 15 samples
(i ¼ 16 , . . . , 30). Here, N(μ, σ 2) is the normal distribution with mean μ and standard
deviation σ (i.e., variance σ 2). The 30 2 matrix in part E shows two environment
types of the 30 samples. The hth column (h ¼ 1 or 2) of the matrix corresponds to
the hth environment type. For the 10 samples (i ¼ 1 , . . . , 5 and 16 , . . . , 20), a
small value (0.3) is intentionally added to yi to reflect putative environmental
effects. According to whether or not a sample is associated with an environment
type, each element of the matrix in part E takes a value of 0 or 1, which are
respectively colored in white and gray. In this example, we assume that the
10 samples (i ¼ 1 , . . . , 5 and 16 , . . . , 20) have environment type E1, and the
other samples have type E2. Part G shows the genotypes of the samples. The
element of the matrix in the ith row and jth column (zij) shows the genotype of
the jth genetic marker of the ith sample. The genotype of the sample is represented
using multiple colors. This example, which assumes that each genetic marker has
two genotypes, uses two colors (white and gray) to indicate the genotypes of the
samples. If a genetic marker has more than two genotypes, additional colors are
used. For example, white, light gray, and dark gray would be used for a genetic
marker with three genotypes. The 30 50 matrix in part G shows the genotypes of
the 30 samples at the 50 genetic markers (G1 to G50). The genotypes are randomly
generated except for the 10th and 40th genetic marker (G10 and G40), so that these
two genetic markers are associated with the distribution of the trait values (the
samples with higher trait values have the genotype in gray with high probability).
3 Derivation of Linear Models for Quantitative Traits by Bayesian Estimation. . . 25
8.0
P E G
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
1
2
3
4
5
6
7
8
9
0.75 0.25
1
2
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
Fig. 3.1 A dataset example for genomic selection
3.3 Prediction Models
One goal of the GS is to predict the trait value of a sample using the obtained
environment types and genotypes; for this purpose, a prediction model that can
output prediction values for the trait is constructed.
3.3.1 Linear Models
One simple format of the prediction model for the target values (i.e., trait values) is
a summation of the effects by the multiple factors (e.g., environmental factors,
genetic factors, and interactions among them). A model with this format is referred
to as a linear model. Here, note that a matrix, A0 , which is also denoted by AT, is the
transpose of a matrix, A, i.e., a matrix whose rows and columns are exchanged.
Note also that variables in bold fonts will hereafter refer to vectors and matrices.
A linear model for the target values of N samples is given as follows:
XM
y ¼ Xβ þ j¼1
Zj uj þ e: ð3:1Þ
Here, y ¼ (y1, . . . , yN)0 represents the target values of the N samples. Xβ is the
term
X M representing the effects by the environmental factors (environment types),
Z u are the terms representing the summation of the effects by the genetic
j¼1 j j
factors (marker genotypes), and e ¼ (e1, . . . , eN)0 are the random residuals.
The target value of a sample is made up of two constituent values, one calculated
as the summation of the values by fixed effects and the other as the summation of
the values by random effects. The linear model given by Eq. 3.1 assumes that the
target value of a sample can be explained by the summation of the values from these
two types of effects, fixed effects and random effects. The reason why this model is
referred to as a linear mixed model X is that these two types of effects are included in
M
it. In relation to Eq. 3.1, Xβ and Z u , respectively, correspond to the fixed
j¼1 j j
effects and the random effects. We assume that the dimensions of X are N L and
that X has full column rank, i.e., rank (X) ¼ L. e ¼ (e1, . . . , eN)0 are the random
residuals that could not be explained by those effects. Letting σ 2e be the variance of
the random residuals, the target values follow the normal distributions as follows:
XM
yj β, u1 , . . . , uM , σ 2e NðXβ þ j¼1
Zj uj , Rσ 2e Þ, ð3:2Þ
where R is a known matrix of N N dimensions. When we assume that the total

samples are divided into several groups according to their observed attribute values,
we can intuitively consider that the fixed effects represent the global baseline value
and the intergroup deviations, while the random effects correspond to the
intragroup deviations.
In relation to the linear models for the trait values, typically and naturally,
environmental factors constitute the fixed effects, while the genetic factors consti-
tute the random effects for the samples in the same environment group. In Eq. 3.1,
β ¼ (β0, β1, . . . , βL)0 represents the effects of the environment types where β0 is a
variable for the global baseline and βh is the effect of the hth environment type
(h ¼ 1 , . . . , L ). The ith row of X is the vector of a dummy value for the global
baseline and the L observed environment types of the ith sample (1, xi1, . . . , xiL).
Here, xih is set to 1 if the ith sample has the hth environment type, and otherwise it is
0 (h ¼ 1 , . . . , L). The ith row of the product Xβ gives the portion of the trait value
that is explained by the environmental
XM 0 factors, β0 + β1xi1 þ . . . þ βLxiL. With
respect to Z u
j¼1 j j
, u j ¼ u j1 ; . . . ; u jqj are the effects of the genotypes of the
jth genetic marker (qj is the number of the genotypes of the jth genetic marker), and
the ith row of Zj shows the observed genotypes of the ith sample, ðzji1 , . . . , zjiqj Þ.
Here, zjit is set to 1 if the ith sample has the tth genotype for the jth genetic marker,
and otherwise it is 0 (t ¼ 1 , . . . , qj). The ith row of the product Zjuj gives the
portion of the trait value that is explained by the jth genetic marker,
uj1 zji1 þ þ ujqj zjiqj . X and Zj are referred to as design matrices. Note that the
elements of X and Zj are constants that indicate whether the samples have specific
conditions (e.g., environment types and marker genotypes) in the given dataset. On
the other hand, the elements of β and uj are variables that must be determined so that
the resulting linear model fits the dataset well and the trait values are predicted with
high accuracy. Using the variables above, Eq. 3.1 can be rewritten as follows:
0 1 0 10 1
y1 1 x11 x1L β0
B y2 C B 1 x21 ⋮ x2L CB β1 C
B C
@⋮A¼@ ⋮ ⋮ ⋮ ⋮
A@
⋮
A
yN 1 x xNL β
0N1 1L ð3:3Þ
zj11 zj1qj 0 1 0 e1
1
u
P M B zj21 ⋮ zj2qj C j1
þ j¼1 B C@ ⋮ A þ B @
e2 C
A:
@ ⋮ ⋮ ⋮ A ujq ⋮
zjN1 zjNqj j
eN
3.3.2 One-Marker Model
We first consider a simple example where no environmental factors (L ¼ 0) and

only a single genetic marker (M ¼ 1) are included as follows:
y ¼ Xβ þ Z1 u1 þ e, ð3:4Þ
0
1 0
where y ¼ (y1, . . . , yN)0 , X ¼ (1, . . . , 1)0 , β ¼ (β0)0 , and Z1 ¼ or . Here,
0 1
we consider that the genetic marker has two genotypes γ 1 and γ 2 (e.g., homozygous
for an allele and heterozygous for two alleles), and a sample takes one of the two
genotypes at the genetic marker. Then, the ith row vector of Z1 is one of (1, 0) and
(0, 1), respectively, if its genotype is γ 1 and γ 2. In Eq. 3.4, u1 ¼ (u11, u12)0 indicates
the effects of those two genotypes of the genetic marker. e ¼ (e1, . . . , eN)0 are the
random residuals. Then Eq. 3.4 can be rewritten as follows:
0 1 0 1 0 1 0 1
y1 1 1 0 e1
B⋮C B ⋮C B or C u11 B ⋮C
@ A¼@ Aðβ0 Þ þ @ A þ@ A: ð3:5Þ
⋮ ⋮ 0 1 u12 ⋮
yN 1 ⋮ eN
Here, yi ¼ β0 + u11 + ei if its genotype is γ 1, and yi ¼ β0 + u12 + ei if γ 2. Letting

a0 ¼ β0 + u11 and a1 ¼ u12 u11, we have:
yi ¼ a0 þ a1 g1i þ ei , ð3:6Þ
where g1i ¼ 1 if the genotype of the genetic marker in the ith sample is γ 2, and
otherwise 0 (i ¼ 1 , . . . , N ).
Note that the number of the genotypes is not restricted to two. If the genetic
marker has a total of q genotypes, the size of the design matrix Z1 is N q, and the
tth column of its ith row vector is set to 1 if the ith sample has the tth genotype at
this genetic marker and otherwise is set to 0. u1 has q elements (u11, . . . , u1q).
3.3.3 Two-Marker Model Without Interactions
In a similar way as for the one-marker model, we can consider another example
where no environmental factors (L ¼ 0) and two markers (M ¼ 2) are included as
follows:
y ¼ Xβ þ Z1 u1 þ Z2 u2 þ e, ð3:7Þ
where y ¼ (y1, . . . , yN)0 , X ¼ (1, . . . , 1)0 , and β ¼ (β0)0 . The terms Z1u1 and Z2u2,
respectively, represent the effects by the two genetic markers. e are the random
residuals. Here, we assume that the two genetic markers are mutually independent
and their effects contribute to the trait value in an additive manner. There are no
epistatic effects caused by their interactions. We also assume that the genetic
markers, respectively, have two genotypes (γ 11 and γ 12 for one genetic marker
while γ 21 and γ 22 for the other). Then, Eq. 3.7 can be rewritten as follows:
0 1 0 1 0 1 0 1
y1 1 1 0 1 0
B⋮C B ⋮C B or C u11 B or C u21
@ A¼@ ⋮A 0
ðβ Þ þ @ A þ@ A
⋮ 0 1 u12 0 1 u22
yN 1 ⋮ ⋮
0 1
e1
B C
þ @ ⋮ A: ð3:8Þ
⋮
eN
According to the genotypes of the two genetic markers, which are either
(γ 11, γ 21), (γ 12, γ 21), (γ 11, γ 22), or (γ 12, γ 22), the trait value of a sample can be written
as follows:
yi ¼ β0 þ u11 þ u21 þ ei , ð3:9Þ

yi ¼ β0 þ u12 þ u21 þ ei
ð3:10Þ
¼ ðβ0 þ u11 þ u21 Þ þ ðu12 u11 Þ þ ei ,
ð3:11Þ
¼ ðβ0 þ u11 þ u21 Þ þ ðu22 u21 Þ þ ei ,
ð3:12Þ
¼ ðβ0 þ u11 þ u21 Þ þ ðu12 u11 Þ þ ðu22 u21 Þ þ ei :
Letting a0 ¼ β0 + u11 + u21, a1 ¼ u12 u11, and a2 ¼ u22 u21, we have
yi ¼ a0 þ a1 g1i þ a2 g2i þ ei , ð3:13Þ
where gji ¼ 1 if the genotype of the ith sample is γ j2, and otherwise 0 ( j ¼ 1
or 2, i ¼ 1 , . . . , N ).
3.3.4 Two-Marker Model with Interactions
We can include the effects by the interactions among the markers, by adding a term
to Eq. 3.7 as follows:
y ¼ Xβ þ Z1 u1 þ Z2 u2 þ Wρ þ e, ð3:14Þ
where W is the design matrix for the interactions among the pairs of the genetic
markers and ρ is the vector of the effects by these interactions. Each column of W
corresponds to a pair of genotypes of the two genetic markers. When the two
genetic markers have, respectively, two genotypes (γ 11/γ 12 and γ 21/γ 22) as in
Eq. 3.7, an example of Wρ can be written as follows:
0 10 1
ðγ 11 γ 21 Þ1 ðγ 11 γ 22 Þ1 ðγ 12 γ 21 Þ1 ðγ 12 γ 22 Þ1 ρ11
B ðγ 11 γ 21 Þ2 ðγ 11 γ 22 Þ2 ðγ 12 γ 21 Þ2 ðγ 12 γ 22 Þ2 C B C
Wρ ¼ B CB ρ12 C:
@ ⋮ ⋮ ⋮ ⋮ A @ ρ21 A
ðγ 11 γ 21 ÞN ðγ 11 γ 22 ÞN ðγ 12 γ 21 ÞN ðγ 12 γ 22 ÞN ρ22
ð3:15Þ
Here, the columns of W correspond to the combinations of the genotypes of the

genetic markers. In this matrix, ðγ 1t1 γ 2t2 Þi , which is for the ith sample, is 1 if the
first genetic marker has the t1th genotype and also the second genetic marker has
the t2th genotype, and otherwise it is 0 (t1 , t2 ¼ 1 or 2 in this example). The elements
of ρ ¼ (ρ11, ρ12, ρ21, ρ22)0 are the effects by the combinations of the genetic markers.
In a similar way, we can arbitrarily include the additional terms in Eq. 3.7; for
example, we can include the combinations between an environmental factor and a
genetic factor, in addition to the combinations among genetic factors as given by
Eq. 3.14. Both the terms for the genetic factors and the terms for their interactions
are given by the products of the design matrix and the vector of the effects.
Therefore, they cannot be distinguished from each other merely by using the
numerical dataset described in the model: some interpretations specific to the
context are required for the modeling. In this sense, even with the interactions
among environmental and genetic factors, the values of the target trait can be
described using the linear mixed model given by Eq. 3.1.
3.4 Decomposition of Variance and Contributions

of Genetic Markers
The linear model given by Eq. 3.1 is rewritten as follows:

XM
y ¼ Xβ þ j¼1
Zj uj þ e ¼ Xβ þ Z1 u1 þ Z2 u2 þ þ ZM uM þ e: ð3:16Þ
For the jth genetic marker, Zjuj is a vector of the genetic values of the samples.
Letting δji be the value associated with the ith sample, we have:
0 10 1 0δ 1
1 or 0 1 or 0 uj1 j1
B ⋮ ⋮ ⋮ ⋮ CB C Bδ C
B CB uj2 C B j2 C
Zj uj ¼ B C@ ¼B C: ð3:17Þ
@ ⋮ ⋮ ⋮ ⋮ A ⋮A @⋮ A
1 or 0 1 or 0 ujqj δjN
Here, only one element of the row vector of Zj is 1 (the rest of the elements are 0)
because a sample can have exclusively one genotype at the jth genetic marker.
Missing data is not considered. If all the pairs of the terms are not correlated
(independent and additive), i.e., their covariance is zero, the total variance V(y)
can be given by the summation of their variances:
V ðyÞ ¼ V ðXβÞ þ V ðZ1 u1 Þ þ V ðZ2 u2 Þ þ þ V ðZM uM Þ þ V ðeÞ: ð3:18Þ
For the jth genetic marker, V(Zjuj) is the variance of the values, δj1 , δj2 , , δjN
in Eq. 3.17. V(Xβ) ¼ 0, if there X are no environmental effects. Since the total
N
variance is obtained by V ðyÞ ¼ i¼1
ðyi μÞ2 =N (μ is the mean of yi), the
contribution of the jth genetic marker is evaluated by

r j ¼ V Zj uj =V ðyÞ: ð3:19Þ
This ratio shows the proportion of the variance explained by the jth genetic
marker. Note again that we are assuming the markers are not correlated and there
are no genetic interactions, i.e., linkage disequilibrium and epistatic effects are not
assumed.
Since the denominator of Eq. 3.19 is constant in a set of samples, the ratio rj is
determined by the product of Zj and uj. As shown in Eq. 3.17, the former represents
the distribution of the genotypes in the samples, while the latter represents the
effects of the genotypes. Therefore, if a portion of the elements of uj are diverse and
distributed in the samples, V(Zjuj) increases and the ratio has a higher value. On the
other hand, if all the row vectors of Zj are identical or all the elements of uj are
identical, the ratio is 0, showing that there are no differences among the samples in
relation to the genetic marker.
Although the ratio of the variances by Eq. 3.19 provides an index for the
estimation of the effects of a genetic marker, it can fail to detect a genotype with
a strong effect but with a low frequency
in the samples.
To remedy this problem, the
0
variance of the elements of uj ¼ uj1 ; uj2 ; . . . ; ujqj is evaluated as follows:
Xqj 2
vj ¼ k¼1
ujk ujμ =qj , ð3:20Þ
Xqj
where ujμ ¼ u =qj . Once uj is obtained, vj is not dependent on the samples,
k¼1 jk
while rj is dependent on the distribution of the genotypes in the samples.
Even with a low frequency in the samples, a genotype highly correlated to the
trait of interest is important, especially in the selection process for the breeding
purposes. As described in the later sections, the variance given by Eq. 3.20 is one of
the key values for elucidating the factors that contributed to the trait.
3.5 Breeding Values and Heritability
To focus on the effects that are inherited from one generation to the next, a trait
value P (also referred to as a phenotype or a phenotypic value) of a sample can be
decomposed in a symbolic manner as follows:
P ¼ G þ E, ð3:21Þ
where G (representing a genetic value or a genotypic value) is the portion that is

determined by the genetic factors, while E (representing an environmental value or
environmental deviation) is the portion that is not explained by the genetic factors
and is attributed to the nongenetic factors (i.e., environmental factors). Here, P is
the observed value in a sample, and G for the sample is the expected value of the
trait in samples with the same genetic background as the sample. E is the random
residual. The deviation from the mean of the trait values in the samples can be used
for the actual value of P for a sample. The deviation is partitioned into the effects of
the genetic factors and the environmental factors. The portion by the genetic factors
G is further divided as follows:
G ¼ A þ D þ I, ð3:22Þ
where A shows the deviations by the additive genetic effects, D shows the devia-
tions by the dominant genetic effects, and I shows the deviations by the effects of
interactions among genetic markers, also referred to as epistatic effects. Here,
A captures the effects of alleles in an additive manner. D captures the nonlinear
effects (i.e., deviations from the linear effects) caused by the combinations of
alleles (heterozygous genotypes) within a genetic marker, while I captures the
nonlinear effects (i.e., deviations from the linear effects) by the combinations of
alleles or genotypes among multiple genetic markers. D and I, respectively, corre-

spond to the deviations by the intra- and inter-marker interactions among alleles.
Here, interactions among G and E are ignored. From Eqs. 3.21 and 3.22, we have:
P ¼ G þ E ¼ ðA þ D þ I Þ þ E: ð3:23Þ
If there are no interactions among the terms of Eq. 3.23, the variance can be
decomposed as follows:
V ðPÞ ¼ V ðGÞ þ V ðEÞ ¼ V ðAÞ þ V ðDÞ þ V ðI Þ þ V ðEÞ: ð3:24Þ
In the traditional GS, we usually consider that only the additive effect A is
transmitted to the progenies in the succeeding generations and only this effect is
referred to as a breeding value (BV). The pairs of alleles of the heterozygous
genotypes (dominant effects) and the combinatorial patterns of alleles among the
genetic markers (epistatic effects) are not entirely passed on to the offspring, and
therefore the effects that are expected to be the same in the progeny, i.e., the
additive effects, are emphasized.
The ratios of the variance of G (genotypic value) and A (additive genotypic
value) against the variance of P in samples are, respectively, referred to as the
broad-sense heritability and the narrow-sense heritability. The broad-sense herita-
bility is defined as follows:
H 2 ¼ V ðGÞ=V ðPÞ: ð3:25Þ
The narrow-sense heritability is defined as follows:
h2 ¼ V ðAÞ=V ðPÞ: ð3:26Þ
The genetic value of a sample is defined by the phenotypic value that can be
attributed to the genetic factors including nonadditive allele interactions. On the
other hand, the breeding value of a sample is defined by the phenotypic value that
can be attributed to only the additive genetic factors.
In this way, the GS is usually based on the effects of additive genotypic values.
One reason why such a strategy has been adopted might be that the decomposition
of phenotypic variance has been considered in a symbolic and abstract manner as in
Eq. 3.24, where the entity of the genetic factors is actually ambiguous. However,
the datasets obtained by the high-throughput sequencers (also known as the next-
generation sequencers (NGS)) and high-density microarrays provide the genotypes
of the samples at single base-pair resolution across their whole-genome sequences.
Consequently, such high-resolution data make it possible to evaluate the genotypic
values based on the actual variations of DNA sequences, e.g., single nucleotide
variations (SNVs), short insertion and deletions (Indels), and structural variations
such as copy number variations (CNVs) instead of putative genetic factors. Espe-
cially in relation to the SNVs, for example, the types of the alleles of a genetic
marker are restricted to those that are represented symbolically by their nucleobases
(A, C, G, and T), and therefore the dominant effect by a heterozygous genotype at a
position in the DNA sequence will be expected to found again in the progeny.
Actually, such point mutations may alter the characteristics of the proteins that they
correspond to. If the trait is controlled by the point mutations, the problem will be
simple. The inter-marker interactions among plural genetic markers may have the
same characteristics because the probability that we will find the same genotype
patterns in the progeny is not zero even if it is low.
A genetic marker represents a chromosomal region which it belongs to, and its
genotype partially shows the way the region has been inherited from the ancestors.
However, even if the identical genotype at a genetic marker is found in the samples,
the genetic factors which were near to the genetic marker are not always the same in
those samples. It depends on how they are genetically related as determined by the
manner of crossing. The identical allele constituting a genotype at a genetic marker
can be found in the samples when it has been inherited from the independent
ancestors having it or it has been caused by spontaneous mutations (identity by
state (IBS)). The identical allele can be found in the samples also when it has been
inherited from the common ancestors (identity by descent (IBD)). If the founder of
the chromosomal regions can be traced back in the ancestors, detection of the
dominant effects by the accompanying genetic factors can be expected even in
their offspring for such regions. An increase in the number of the genetic markers
makes it possible to capture the genetic factors by the patterns of their genotypes
(haplotypes) instead of the genotype of a single genetic marker; however, improve-
ment of the resolution introduces nonadditive effects into the datasets. Linkage
disequilibrium (LD) among the genetic markers in a chromosome (cis-interactions)
must be taken into consideration in addition to their genetic relationships across
chromosomes (trans-interactions).
3.6 Determination of Model Parameters by Least-Squares

Estimation
The least-squares estimation approach can analytically determine the parameters in

the models. The random residuals, i.e., the deviations of the predicted values from
the observed real values, are minimized in the samples. The resulting models
identify the underlying causal factors in the given datasets and characterize the
target values by using those factors. We consider, for example, a one-marker model
by Eq. 3.6 for a dataset with N ¼ 2 samples and M ¼ 1 genetic marker. When the
target values and the 0/1-encoded genotypes are given, respectively, by y ¼
(y1, y2)0 ¼ (2, 1)0 and t1 ¼ (g11, g12)0 ¼ (1, 0)0 , the relationships among the parame-
ters in the model are given as follows:
2 ¼ a0 þ a1 1 þ e 1
: ð3:27Þ
1 ¼ a0 þ a1 0 þ e 2
Letting the random residuals e ¼ (e1, e2)0 be 0, the equations can be uniquely
solved (a0, a1) ¼ (1, 1). However, if an additional sample is given (y3, g13) ¼
(2.1, 1), for example, the equations cannot be solved under the assumption that
the random residuals are 0. Some portion of the target values must be assigned to
the random residuals under the restriction that they are minimized, 0.1 from y3 to e3
in this case. If any genetic markers in a model are not associated with the target
values, by setting all the parameters except the random residuals to 0 and assigning
all the values to the random residuals, we have a trivial solution y ¼ e. Thus, the
problem can be solved by minimization of the random residuals.
3.6.1 Least-Squares Estimation for One-Marker Model
From Eq. 3.6, we have ei ¼ yi a0 a1g1i. Letting J denote the squared error as
follows, a0 and a1 are determined so that J is minimized:
XN XN
J¼ e2 ¼
i¼1 i i¼1
ðyi a0 a1 g1i Þ2 : ð3:28Þ
XN
Letting ∂J=∂a0 ¼ 2 i¼1
ðyi a0 a1 g1i Þ ¼ 0, we have:
XN y XN g
a0 ¼ i¼1 N
i
a1 i¼1 N
1i
: ð3:29Þ
XN
Similarly, letting ∂J=∂a1 ¼ 2 ðy a0 a1 g1i Þ∂ðyi a0 a1 g1i Þ=∂a1 ¼ 0
i¼1 i
and replacing a0 with Eq. 3.29, we have:
∂J XN n X N y XN g on X N g o
¼2 y i
a 1
1i
g a 1 g 1i
∂a1 i¼1 i i¼1 N i¼1 N 1i 1i i¼1 N
¼ 0:
ð3:30Þ
Manipulating Eq. 3.30, we have:

X X N y XN g XN X N g 2
∂J N
¼ 2 yi i
g1i 1i
a1 g1i 1i
∂a1 i¼1 i¼1 N i¼1 N i¼1 i¼1 N
¼ 0:
ð3:31Þ
Equation 3.31 can be written as ∂J/∂a1 ¼ 2N(σ y1 a1σ 11) ¼ 0. Here, σ y1 is

the covariance between y ¼ (y1, y2, . . . , yN)0 and t1 ¼ (g11, g12, . . . , g1N)0 . σ 11 is the
variance of t1. If σ 11 6¼ 0, we have:
P N n P N yi P N g1i o
σ y1 i¼1 y i i¼1 N t1i i¼1 N =N
a1 ¼ ¼ PN P N g1i : ð3:32Þ
σ 11 2
i¼1 g1i i¼1 N =N
The prediction value for the ith sample is given as follows:
pi ¼ a0 þ a1 g1i : ð3:33Þ
3.6.2 Least-Squares Estimation for Two-Marker Model

Without Interactions
From Eq. 3.13, we have ei ¼ yi a0 a1g1i a2g2i. In a similar way as for the
one-marker model, the squared error J is given as follows and minimized:
XN
J¼ i¼1
ðyi a0 a1 g1i a2 g2i Þ2 : ð3:34Þ
XN
Letting ∂J=∂a0 ¼ 2 i¼1
ðyi a0 a1 g1i a2 g2i Þ ¼ 0, we have:
XN y XN g XN g
a0 ¼ i¼1 N
i
a 1
1i
i¼1 N
a 2
2i
i¼1 N
: ð3:35Þ
Replacing a0 with Eq. 3.35 in Eq. 3.13, we have:

XN y XN g XN g
yi ¼ k
a 1
1k
a 2
2k
þ a1 g1i þ a2 g2i þ ei
k¼1 N
XN y k¼1 N
XN g k¼1 N XN g ð3:36Þ
¼ k¼1 N
k
þ a 1 g 1i k¼1 N
1k
þ a 2 g 2i 2k
k¼1 N
þ ei :
Using Eq. 3.36, Eq. 3.34 can be rewritten as follows:

XN n XN y XN g X N g o2
J¼ i¼1
y i k¼1 N
k
a 1 g 1i 1k
k¼1 N
a 2 g 2i 2k
k¼1 N
:
ð3:37Þ
Then, we have:
∂J X N n XN y XN g X N g
¼ 2 y k
a 1 g 1k
a2 g 2k
∂a1
i¼1
P N g1k
i
o
k¼1 N 1i k¼1 N 2i k¼1 N
g1i
nP N P N yk P N g1k
k¼1 N
¼ 2 y i g 1i
PN P N Ng 2
i¼1 k¼1 k¼1 N
a1 g1i 1k
Pi¼1N Pk¼1 N
N g2k P N g1k o
a2 i¼1
g2i k¼1 N
g1i k¼1 N
:
ð3:38Þ
Therefore, we also have:
∂J
¼ 2N σ y1 a1 σ 11 a2 σ 12 ð3:39Þ
∂a1
and
∂J
¼ 2N σ y2 a1 σ 12 a2 σ 22 : ð3:40Þ
∂a2
Here, σ y1 is the covariance between y and t1 ¼ (g11, g12, . . . , g1N)0 , σ y2 is the

covariance between y and t2 ¼ (g21, g22, . . . , g2N)0 , and σ 11 and σ 22 are the variance
of t1 and t2. σ 12 and σ 21 are the covariance between t1 and t2 (σ 12 ¼ σ 21). Letting
∂J/∂a1 ¼ 0 and ∂J/∂a2 ¼ 0, we have:

a1 σ 11 þ a2 σ 12 ¼ σ y1
: ð3:41Þ
a1 σ 21 þ a2 σ 22 ¼ σ y2
If σ 11 σ 22 σ 212 6¼ 0, we have:
σ y1 σ 22 σ y2 σ 12 σ y2 σ 11 σ y1 σ 12
a1 ¼ and a2 ¼ : ð3:42Þ
σ 11 σ 22 σ 12
2 σ 11 σ 22 σ 212
Equation 3.41 can also be described using matrices as follows:


σ 11 σ 12 a1 σ y1
¼ : ð3:43Þ
σ 21 σ 22 a2 σ y2

σ 11 σ 12
Letting Σ ¼ and its determinant detΣ ¼ σ 11σ 22 – σ 12σ 21, if detΣ 6¼ 0,
σ 21 σ 22
1 σ 22 σ 12
there exists the inverse of Σ, given by Σ ¼ detΣ 1
. Then we have:
σ 21 σ 11

a1 1 σ y1 1 σ 22 σ 12 σ y1
¼Σ ¼ : ð3:44Þ
a2 σ y2 detΣ σ 21 σ 11 σ y2
Equation 3.44 is equivalent to Eq. 3.42.
3.6.3 Number of Parameters
As shown in the examples in the previous sections, the number of parameters that
must be determined increases with the increase in the number of the genetic
markers in the model. Intuitively, N independent equations are required to uniquely
determine N parameters. In the typical datasets we use for the GS and the MAS, the
number of the genetic markers p is much greater than that of the samples N;
therefore, they cannot provide a sufficient number of equations for the parameters
that are associated with the genetic markers. We thus cannot uniquely determine the
parameters. This situation is sometimes referred to as the N p problem.
Let f(gi) ¼ a0 + a1g1i þ . . . þ aMgMi be the formula for the prediction model.
Letting a ¼ (a0, a1, . . . , aM)0 be the vector of the coefficients and gi ¼ (g0, g1i, . . . ,
gMi) be the vector of the observed values (e.g., 0/1-encoded genotypes) for the ith
sample ( f(gi) ¼ gia) with a dummy variable g0 1, we have N predictions for
N samples. Letting G ¼ (g1, g2, . . . , gN)0 , those predictions are written as Ga.
Then, the squared error J is given as follows:
J ¼ ðGa yÞ0 ðGa yÞ: ð3:45Þ
Equations 3.28 and 3.34 in the previous sections can be written in this format.
Here, y0 Ga ¼ (y0 Ga)0 ¼ a0 G0 y because it is a scalar, and we then have:
J ¼ ða0 G0 y0 ÞðGa yÞ ¼ a0 G0 Ga 2a0 G0 y þ y0 y: ð3:46Þ
The derivative of J with respect to a is given as follows:
∂J=∂a ¼ 2G0 Ga 2G0 y: ð3:47Þ

From Eq. 3.47, we have:
G0 Ga ¼ G0 y: ð3:48Þ
This equation is referred to as the normal equation. When G0 G is a regular

matrix, there exists its inverse and a is uniquely determined in an analytical manner
as follows:
a ¼ ðG0 GÞ1 G0 y: ð3:49Þ
However, as noted above, the number of the parameters is greater than that of the
equations given by y ¼ Ga, and hence the parameters cannot be uniquely deter-
mined. Instead of solving the equations in an analytical manner, therefore, methods
based on model fitting to the data are used so that the squared errors are minimized.
Although the number of the parameters is greater than that required to solve the
equations, all the parameters do not contribute to the values of the dependent
variable, y. Therefore, in parallel with determining the values of the parameters,
the selection of the parameters is explicitly and implicitly carried out. As a result,
the parameters for the independent variables that do not contribute to the dependent
variable will converge to zero and be excluded from the prediction model.
3.7 Determination of Model Parameters by Bayesian

Estimation
The Bayesian approach can determine the parameters in the models. As we have
seen in the previous sections, the random residuals associated with the prediction
model can be assumed to follow some distribution, e.g., the normal distribution.
The characteristics of the distribution of the random residuals such as the mean and
the variance are important for evaluation of model fitting to a given dataset rather
than the values assigned to the samples. The random residuals can be thus evaluated
by using the distribution of a random variable. In a similar way, if we represent the
parameters in the prediction model by using the distributions of random variables,
we can intuitively introduce the Bayesian approach into the construction of the
prediction model. The ranges with high probability can estimate the likely values
for the parameters in a stochastic manner. A part of this section in relation to the
Gibbs sampling is based on Wang et al. (1993).
3.7.1 Basic Concepts of Bayes’ Theorem
The joint probability can be written as follows:

PðA; BÞ ¼ PðAjBÞPðBÞ ¼ PðBjAÞPðAÞ: ð3:50Þ
Here, P(AjB) and P(BjA) are the conditional probabilities of A and B, respec-
tively, given B and A. P(A) and P(B) are the marginal probabilities. If P(B) 6¼ 0, we
have:
PðBjAÞPðAÞ
PðAjBÞ ¼ / PðBjAÞPðAÞ: ð3:51Þ
PðBÞ
This relationship is referred to as the Bayes’ theorem and shows that the
probability of A given B is proportional to the product of the probability of A and
the probability of B given A. Using this relationship, we can estimate the posterior
probability P(AjB) from the prior probability P(A). The conditional probability
P(BjA) shows the probability of obtaining B under the condition A, which is referred
to as the likelihood.
In Eq. 3.50, the probabilities are assigned to discrete events represented by A and
B. This concept can be extended to the continuous distributions of a parameter θ as
follows:
π ðθjDÞ / f ðDjθÞπ ðθÞ: ð3:52Þ
Here, a parameter θ is expressed by using a distribution of probability instead of

a single value. The value of θ with a high probability is likely to be the true value
of θ. In Eq. 3.52, the product of the prior distribution π(θ) and the likelihood of
obtaining the dataset D under the condition θ, i.e., f(Djθ), provide the evaluation for
the posterior distribution π(θjD).
3.7.2 Estimation of Parameters of Normal Distributions
Here, we consider the case that the dataset D consists of N observations of xi

(i ¼ 1 , 2 , . . . , N). We also assume the condition that the distribution of xi follows
a normal distribution with the mean μ and the variance σ 2. Here, μ and σ 2 are not
known, and they are the targets of the prediction. If the N observations are
independent of each other, the likelihood of obtaining the dataset D is given by
the product of the probability density functions as follows:
!
YN 1 ðxi μÞ2
f ðDjμ, σ 2 Þ ¼ pffiffiffiffiffiffiffiffiffiffi exp : ð3:53Þ
i¼1
2πσ 2 2σ 2
If we set values for μ and σ 2, the probability of the event that we obtained the
dataset D is given by f(Djμ, σ 2). This value is referred to as the likelihood under the
condition that the values follow the normal distribution. The plausible values for μ
and σ 2 can be determined so that the likelihood is maximized.
3.7.2.1 Estimation of Mean with Known Variance
Suppose we are estimating μ under the condition that σ 2 is known. As the prior
probability distribution for μ, we can use a normal distribution with a mean μ0 and a
variance σ 20 as follows:
!
1 ðμ μ 0 Þ2
π ðμÞ ¼ pffiffiffiffiffiffiffiffiffi2ffi exp : ð3:54Þ
2πσ 0 2σ 20
If σ 20 is a very large value, the distribution is near to a flat distribution, showing

that there is no prior knowledge. Then, we have:
! !
YN 1 ðxi μÞ2 1 ðμ μ 0 Þ2
π ðμjDÞ / pffiffiffiffiffiffiffiffiffiffi exp pffiffiffiffiffiffiffiffiffi2ffi exp : ð3:55Þ
i¼1
2πσ 2 2σ 2 2πσ 0 2σ 20
2
Manipulating Eq. 3.55 by using ðxi μÞ2 ¼ xi x þ x μ where
XN
x ¼ x , we have:
i¼1 i
2 !
Nσ 2 x þμ0 σ 2 σ 2 σ 20
πðμjDÞ / exp μ 0 2 2 : ð3:56Þ
Nσ 0 þ σ 2 Nσ 20 þ σ 2
Letting
Nσ 20 x þ μ0 σ 2 σ 2 σ 20
μ1 ¼ and σ 2
¼ , ð3:57Þ
Nσ 20 þ σ 2 1
Nσ 20 þ σ 2
we have:
!
ðμ μ 1 Þ2
π ðμjDÞ / exp : ð3:58Þ
2σ 21
This shows that the posterior probability distribution is again a normal distribu-
tion and has its maximum value at μ1. Therefore, μ1 is adopted for the estimated
value of μ. This method is referred to as the maximum a posteriori probability
(MAP) estimation. Letting τ ¼ 1/σ 2 and τ0 ¼ 1=σ 20 in Eq. 3.57, we have:
1 1 1 τN τ0
τ1 ¼ ¼ N þ 2 ¼ τN þ τ0 and μ1 ¼ x þ μ : ð3:59Þ
σ 21 σ 2 σ0 τN þ τ0 τN þ τ0 0
τ is the inverse of the variance and is referred to as the precision. Equation 3.59
shows that the precision is improved by the summation of the precision of the
samples observed (τN > 0) while the mean is updated to the weighted average of the
mean in the samples observed (x) and that of the prior probability distribution (μ0).
3.7.2.2 Estimation of Variance with Known Mean
On the other hand, suppose we are estimating σ 2 (σ 2 6¼ 0) under the condition that μ
is known. From Eq. 3.53, we have:
PN !
1 ð x i μ Þ 2
f Djσ 2 / exp i¼1
: ð3:60Þ
ðσ 2 ÞN=2 2σ 2
As the prior distribution for σ 2, we can use a distribution with parameters ν0 (the
number of chi-squared degrees of freedom) and s20 (the scaling parameter) as
follows:

1 νs20
π σ2 / exp 2 : ð3:61Þ
ðσ 2 Þ1þν0 =2 2σ
Then the posterior probability distribution is as follows:

PN !
1 ðxi μ Þ2 1 ν0 s20
π σ jD /
2
exp i¼1
exp : ð3:62Þ
ðσ 2 ÞN=2 2σ 2 ðσ 2 Þ1þν=2 2
Manipulating Eq. 3.62, we have:

0 PN 1
ν0 s20 þ ðxi μÞ2
B ðν0 þ NÞ
i¼1
1 ν0 þN C
πðσ 2 jDÞ / exp@ A: ð3:63Þ
ðσ 2 Þ1þðν0 þNÞ=2 2σ 2
Letting
PN
ν0 s20 þ ðxi μ Þ2
ν1 ¼ ν0 þ N and s21 ¼ i¼1
, ð3:64Þ
ν0 þ N
we have:

1 ν1 s21
π σ 2 jD / exp : ð3:65Þ
ðσ 2 Þ1þν1 =2 2σ 2
This shows that the posterior probability is the same distribution as the prior
probability distribution. This distribution is referred to as the scaled inverse
chi-squared distribution, and the probability density function is given as follows:

exp νs2x
2
ν=2
ðs ν=2Þ
2
f ðx; ν, s2 Þ ¼ : ð3:66Þ
Γðν=2Þ x1þν=2
Here, Γ(x) is the gamma function and defined for positive real numbers as
follows:
Z 1
ΓðxÞ ¼ tx1 et dt, ðx > 0Þ, ð3:67Þ
0
where Γ(1) ¼ 1 and Γ(x þ 1) ¼ xΓ(x) for positive real numbers. Γ(x) can be consid-
ered to be the extension of the factorial of natural numbers, f(n) ¼ n!.
The scaled inverse chi-squared distribution is derived from the normal distribu-
tion. When the values of the random variable x are independent and followX N(0, s2),
ν
summations of the squares of the values of the random variable, z ¼ 2
x2 ,
i¼1 i
follow the chi-squared distribution with ν degrees of freedom. The probability
density function of the chi-squared distribution (x 0) is given as follows:

ð1=2Þν=2 exp 2x
f ðx; νÞ ¼ : ð3:68Þ
Γðν=2Þ x1ν=2
Letting y ¼ νs2/x (i.e., x ¼ νs2/y), the probability density function of y is given as

follows:
2
ν=2 exp νs
dx ðs ν=2Þ
2
gðy; ν, s2 Þ ¼ f ðx; νÞ ¼
2y
: ð3:69Þ
dy Γðν=2Þ y1þν=2
This distribution is referred to as the scaled inverse chi-squared distribution.

Here, s2 is referred to as a scaling parameter. This shows that the values of νs2/z2
X ν ðxi =sÞ2
follow the scaled inverse chi-squared distribution. By νs2 =z2 ¼ 1= ,
i¼1 ν
2
this implies that if the values of x follow N(0, s ), the inverses of the variances of the
2 2
values of x follow
X N the scaled2 inverse chi-squared distribution, Scaled-inv-χ (ν, s ).
Letting s ¼
2
ðx μÞ in Eq. 3.64, we have:
i¼1 i
N ν0
ν1 ¼ ν0 þ N and s21 ¼ s2 þ s2 : ð3:70Þ
ν0 þ N ν0 þ N 0
Equation 3.70 shows that the degrees of freedom are improved by the number of
the samples observed (N > 0), while the scaling parameter is updated to the
weighted average of the scaling parameter in the samples observed (s2) and that
of the prior probability distribution (s20 ).
3.7.2.3 Natural Conjugate Prior Probability Distributions
When the posterior probability distribution (or simply the posterior) is in the same
family as the corresponding prior probability distribution (or simply the prior) in
relation to a given likelihood function, they are referred to as conjugate distribu-
tions. Such a prior probability distribution is referred to as a (natural) conjugate
prior probability distribution. For example, the probability distribution of a normal
distribution has the maximum at the mean, and the estimator can be easily obtained
from the posterior probability distribution. However, this is not the case with the
arbitrary distributions. Generally, the characteristics of the posterior probability
distributions are not clear if the prior and posterior probability distributions are not
conjugate. When the posterior probability distribution is not well known, numerical
integration is required to obtain its characteristics (e.g., the mean and the variance).
However, this is not always easy for the unknown distributions or complex distri-
butions (e.g., high-dimensional distributions with multiple variables).
3.7.3 Estimation of Linear Mixed Models by Sampling
One method for obtaining the characteristics of the complex posterior probability
distributions is to approximate the distribution by sampling the distribution. This
sampling procedure is also sometimes referred to as simulation. Markov chain
Monte Carlo (MCMC) methods are a class of methods that simulate the distribu-
tions by iterated sampling. Monte Carlo (MC) methods are a class of methods for
numerical analysis using random numbers. Markov chain Monte Carlo methods are
categorized into Monte Carlo methods and are a class of algorithms that carry out
the sampling of the probability distributions based on the Markov chain that has the
target distribution as its equilibrium distribution. Here, a Markov chain is a sto-
chastic process that has the Markov property (also known as the property of
memorylessness), i.e., the conditional probability distribution of the future can be
determined only by the current state. The Gibbs sampling method is one of the
MCMC methods and is applicable when the conditional distribution of each
variable (full conditional distribution) is given.
3.7.3.1 Parameters of Linear Mixed Models for Bayesian Estimation
Let us consider the linear model given by Eq. 3.1 again:

XM
y ¼ Xβ þ j¼1
Zj uj þ e: ð3:71Þ

2 model,2 θ ¼2 β; u; v; σ e , by Bayes’ estimation.
2
We obtain the parameters of this
Here, u ¼ (u1, . . . , uM) and v ¼ σ 1 ; . . . ; σ M . σ j ( j ¼ 1 , . . . , M ) is the variance of

uj ¼ uj1 ; . . . ; ujqj , and σ 2e is the variance of e ¼ (e1, . . . , eN). All the parameters
are unknown. Note that σ 2j is the variance of the effects of the genotypes, not the
variance of the genotypic values. If we obtain the estimation of uj, however, we can
derive the variance of the genotypic values. In the succeeding sections, the prob-
ability distributions required for Gibbs sampling, i.e., the prior probability distri-
butions (3.7.3.2), the joint posterior probability distributions (3.7.3.3), and the
conditional probability distributions for the parameters (3.7.3.4–3.7.3.6), will be
introduced in detail. Then, an implementation of Gibbs sampling using those
distributions will be introduced (3.7.3.7). Finally, the results of estimation of the
linear mixed model for the example given in Fig. 3.1 will be presented.
3.7.3.2 Prior Probability Distributions
Consider the case that the prior probability distributions for the parameters are as
follows. For the fixed effects, we will assume a flat prior probability distribution for
the naive prior probability distribution, which represents the case that we are not
given enough knowledge, as follows:
π ðβÞ / constant: ð3:72Þ
When qj is the number of the genotypes for the jth genetic marker ( j ¼ 1 , . . . , M)
and uj follows a qj-dimensional multivariate normal distribution,
uj j Gj , σ 2j N qj ð0, Gj σ 2j Þ, where Gj is a known matrix indicating the relationships
among uj1 , . . . , ujqj (the effects by the genotypes), the prior probability distribution is
given as follows:
1
1 1 0
π uj jGj ; σ 2j / qffiffiffiffiffiqj exp uj Gj σ j
2
uj
σ 2j 2
! ð3:73Þ
1 1 0 1
/ qffiffiffiffiffiqj exp 2 uj Gj uj :
σ2 2σ j
j
XM 0 XM
Letting s2e ¼ y Xβ j¼1
Z j u j y Xβ j¼1
Z j u j =N and
s2j¼ u0j G1
j uj =qj ,
the prior probability distribution for the variances is given using
the scaled inverse chi-squared distributions as follows:

1 ν e s2
π σ 2e jνe , s2e / 1þν =2 exp 2e ð3:74Þ
σ 2e e 2σ e
and
!
1 νj s2j
π σ 2j jνj , s2j / 1þν =2 exp 2 : ð3:75Þ
j 2σ j
σ 2j
Letting νe ¼ 0 and νj ¼ 0 in Eqs. 3.74 and 3.75 for the naive prior probability
distributions, we have:
1
π σ 2e / 2 ð3:76Þ
σe
and
1
π σ 2j / 2 : ð3:77Þ
σj

The joint prior density of θ ¼ β; u; v; σ 2e is given by the product of the density
given by Eqs. 3.72, 3.73, 3.76, and 3.77 as follows:
YM YM
π ð θÞ ¼ π ð βÞ j¼1
π u j jG j , σ j
2
j¼1
π σ 2j π σ 2e : ð3:78Þ
3.7.3.3 Joint Posterior Probability Distributions
If we assume that R ¼ I (the identity matrix) in Eq. 3.2, i.e., the samples are
independent to each other and the random residuals of all the samples have the
identical variance, the likelihood associated with the N samples is given as follows:

YN 1 1 XM 0 XM
πðyjθÞ ¼ pffiffiffiffiffi exp 2 y Xβ Zu y Xβ Zu :
i¼1 σ 2e 2σ e j¼1 j j j¼1 j j
ð3:79Þ
By using Eqs. 3.78 and 3.79, the joint posterior probability distribution is given
as follows:
πðθjyÞ ¼ πðβ, u, v, σ 2e jyÞ / πðyjθÞ πðθÞ

1 1 XM 0 XM
/ exp y Xβ Z
j¼1 j j
u y Xβ Z
j¼1 j j
u
ðσ 2e ÞN=2 2σ 2e
8 9
!>
YM > < 1 1 = YM 1 1
constant j¼1 q =2 exp 2 uj 0 G1 uj j¼1 2 2
>
: σ2 j 2σ j
j
>
; σ j σ e
j

1 1 XM 0 XM
/ N=2þ1 exp 2 y Xβ Zu
j¼1 j j
y Xβ Zu
j¼1 j j
σ 2e 2σ e
8 9
> !>
YM < 1 1 =
j¼1 q =2þ1 exp 2 uj 0 G1 u j : ð3:80Þ
>
: σ2 j 2σ j j
>
;
j

Here, θ ¼ β; u; v; σ 2e . y are the trait values observed.
3.7.3.4 Conditional Probability Distribution for β
By setting a variable unknown under the conditions that all other variables are
known, as seen in the simple examples, we obtain the conditional posterior prob-
ability distributions. Letting u, v, σ 2e , and y be known, we have:

1 XM 0 XM
π βju, v, σ 2e , y / exp 2 Xβ y þ Z u
j¼1 j j
Xβ y þ Z u
j¼1 j j
:
2σ e
ð3:81Þ
XM XM
Note that y Xβ Z u ¼ Xβ y þ
j¼1 j j j¼1
Zj uj . Then, we have:
XM XM
1
Xβ y þ j¼1
Z j u j ¼ XX Xβ y þ j¼1
Z j u j
n XM o
¼ X β X1 y j¼1
Z j u j : ð3:82Þ
Using (AB)0 ¼ B0 A0 for matrices A and B, the argument of the exponential

function in Eq. 3.81 can be rewritten as follows:

1 XM 0 XM
π βju, v, σ 2e , y / exp 2 Xβ y þ Zu
j¼1 j j
Xβ y þ Zu
j¼1 j j
2σ e
oi
1 h n XM oi0 h n XM
¼ exp 2 X β X1 y j¼1
Z j u j X β X 1
y j¼1
Z j u j
2σ e
o
1 n XM o0 n XM
¼ exp 2 β X1 y Z
j¼1 j j
u X 0
X β X 1
y Z
j¼1 j j
u
2σ e
n XM o0 X0 X n XM o
1
¼ exp β X1 y Z j u j β X 1
y Z j u j :
2 j¼1 σ 2e j¼1
ð3:83Þ
1 1 1
X M ¼ B A
Using (AB) for matrices
1 0
X M A and B, we have
β ¼ X1 y Z u
j¼1 j j
¼ ð X 0
X Þ X y Z u
j¼1 j j
. Then we have:

1 0
π βju, v, σ 2e , y / exp β β Σ1
β β β ,
2
1
where Σβ ¼ ðX0 XÞ σ 2e (i.e., Σ1 0
β ¼ X X=σ e ). Thus, we obtain:
2

β j u, v, σ 2e , y Nðβ, Σβ Þ: ð3:84Þ
3.7.3.5 Conditional Probability Distribution for uj
In a similar way, we consider the case in which all variables except uj are known
( j ¼ 1, . . . , M ).X
Letting uj be the vector obtained by removing uj from u and
M
pj ¼ y Xβ Z u , we have:
k¼1, k6¼j k k
πðuj jβ, uj , v, σ 2e , yÞ

!
1 0 1 0 1
/ exp 2 Zj uj pj Zj uj pj exp 2 uj Gj uj
2σ e 2σ j
!
1 0 1 0 1
1 0 1
¼ exp 2 uj Zj pj Zj Zj uj Zj pj exp 2 uj Gj uj
2σ e 2σ j
( )
1 1
0
0

1
1 0 1
¼ exp 2 uj Zj pj Zj Zj uj Zj pj 2 uj Gj uj : ð3:85Þ
2σ e 2σ j
Here, note that Zj0 Zj and G1 j are symmetric (i.e., Gj is also symmetric), and
letting Zj ¼ (z1, z2, , zn) be a vector of row vectors, the (m, n)th and (n, m)th
element of Zj0 Zj are, respectively, (Zj0 Zj)mn ¼ zn0 zm and (Zj0 Zj)nm ¼ zm0 zn. G1
j is a
multiple of a covariance matrix. Here, (m, n)th element is in the mth row and in the
nth column. Generally, in relation to a summation of two quadratic forms, we have:

0
u u Z0 Z u u þ u0 Gu
n o0 n o
1 1
¼ u ðZ0 Z þ GÞ Z0 Zu ðZ0 Z þ GÞ u ðZ0 Z þ GÞ Z0 Zu
1
1
þ u0 ðZ0 ZÞ þ G1 u: ð3:86Þ
Thus, we have:
πðuj jβ, uj , v, σ 2e , yÞ

( )
1 0 1 0 1
¼ exp 2 uj Z1 0 1
j pj Z j Z j uj Z j pj uj G j uj
2σ e 2σ 2j
8 ! 0 !1 11 9
< 1 0 1 =
0 Zj 0 Zj G1 1 0 Z Z G1
j
uj uj uj @ A uj
j j j
¼ exp uj uj þ 2 þ
: 2 σe 2 σj 2 σe2 σj2 ;
( ! )
1 0 Zj 0 Zj G1 j
/ exp uj uj þ 2 uj uj : ð3:87Þ
2 σ 2e σj
Here,
!
1 1
Zj 0 Zj Gj
uj ¼ þ 2 Zj 0 Zj Z1
j pj
σ 2e σj
!
1 1
0
Zj Zj Gj XM
¼ þ 2 Zj 0 y Xβ Z k u k , ð3:88Þ
σe 2 σj k¼1, k6¼j
and
0 1
! !1
Zj Zj Gj 0 σ 2e
Σ1 ¼ þ 2 and Σj ¼ Zj Zj þ G1 σ 2e : ð3:89Þ
j
σ 2e σj j
σ 2j
Thus, we have:
uj j β, uj , v, σ 2e , y Nðuj , Σj Þ: ð3:90Þ

3.7.3.6 Conditional Probability Distributions for σ 2j and σ 2e
We then consider the case in which all variables except σ 2j are known ( j ¼ 1, . . . , M).
Letting vj be the vector obtained by removing vj from v, we have:
!
1 1 0
πðσ 2j jβ, u, vj , yÞ / exp 2 uj G1
j uj : ð3:91Þ
ðσ 2j Þqj =2þ1 2σ j
Thus, we have:
0
σ 2j j β, u, v, y Scaledinvχ 2 ðqj , uj G1
j uj qj Þ: ð3:92Þ
In a similar way, considering the case in which all variables except σ 2e are known,
we have:
πðσ 2e jβ, u, v, yÞ
1 1 XM 0 XM
/ exp y Xβ Z j u j y Xβ Z j u j :
ðσ 2e ÞN=2þ1 2σ 2e j¼1 j¼1
ð3:93Þ
Thus, we have:
σ 2e j β, u, v, y Scaledinvχ 2 ðN, s2e Þ, ð3:94Þ

PM 0 PM
where s2e ¼ ðy Xβ j¼1 Zj uj Þ ðy Xβ j¼1 Zj uj Þ=N
3.7.3.7 Gibbs Sampling
Based on the probability distributions in the preceding sections, we can implement

a program that carries out the Gibbs sampling procedure. By Eqs. 3.84, 3.87, 3.92,
and 3.94, we have the conditional probability distributions for β, uj, σ 2j , and σ 2e .
Setting a portion of the parameters in Eq. 3.78 as known, we obtain the multivariate
normal distributions for the coefficients for the fixed and random effects (β and uj)
and the scaled inverse chi-squared distributions for the variances of their elements
(σ 2j and σ 2e ).
We can obtain the values by sampling (simulating) using the conditional prob-
ability distributions. The values are obtained as the random numbers generated by
the given normal distribution and the scaled inverse chi-squared distributions.
Starting with arbitrary initial values for the parameters (for the round 0), the
following steps are iterated T rounds until the values converge ( j ¼ 1 , . . . , M ):
1. Update β by Eq. 3.84 under the condition that u, v, and σ 2e are known.
2. Update uj by Eq. 3.87 under the condition that β, uj, v, and σ 2e are known.
3. Update σ 2j by Eq. 3.92 under the condition that β, u, vj, and σ 2e are known.
4. Update σ 2e by Eq. 3.94 under the condition that β, u, and v are known.
Here, uj and vj are, respectively, the vector obtained by removing the jth
element uj from u, and the vector obtained by removing the jth element vj from v.
Accumulating the values sampled and obtaining the frequencies of those values (the
histograms of the frequency distributions), the posterior probability distributions of
the variables are estimated.
In the rounds above, normally distributed random numbers are obtained by
using, for example, the Box-Muller method (Box and Muller 1958), which trans-
forms uniformly distributed random numbers to normally distributed random num-
bers. Random numbers following the scaled inverse chi-squared distribution are
obtained by using the gamma distribution. When X ~ Scaled-inv-χ 2(ν, τ2), we can
generate the random numbers that follow the distribution using the inverse gamma
distribution and the gamma distribution as follows:

X InvGammaðν=2, ντ2 =2Þ and 1=X Gamma ν=2, 2=ðντ2 Þ ð3:95Þ
The random numbers following the gamma distribution can be procedurally

generated (Tanizaki 2008).
To avoid dependency on the initial values, the initial rounds are discarded as the
burn-in period. Also, to avoid dependency among the nearby rounds, the values are
adopted to construct the distributions (i.e., frequencies of the values sampled) at
intervals (thinning). Using the values obtained, the target distributions are
estimated.
3.7.3.8 An Example of Parameter Estimation by Gibbs Sampling
Consider the case of a linear mixed model with five genetic markers (G10, G20,
G30, G40, and G50) with two environmental factors (E1 and E2) as follows:
0 1
β0
@ A u11 u21 u31 u41
y ¼ X β1 þ Z10 þ Z20 þ Z30 þ Z40
u12 u22 u32 u42
β2

u51
þ Z50 þ e: ð3:96Þ
u52
The results of the estimation of the parameters by the Gibbs sampling explained
in the preceding sections can be visualized as in Fig. 3.2. The line graphs in the right
panels present the values of the parameters sampled (T ¼ 1000 rounds), showing
Fig. 3.2 Estimation of the parameters by Gibbs sampling

that the values converge after the initial disturbance. The histograms in the left
panels present the estimated distributions (posterior probability distributions) of the
parameters by using the sampled values obtained in the right panels. For each of uj
( j ¼ 1, . . . , 5), the red and blue bars, respectively, correspond to uj1 and uj2. The red
and blue bars, respectively, correspond to β1 and β2 for β. The first hundred rounds
were discarded as the burn-in period (i.e., the size of the burn-in period is 100).
Thinning was not carried out in this example, and all the values after the burn-in
period were adopted to construct the distributions.
Among the five genetic markers, the 10th genetic marker (G10) has the differ-
ence between the effects of the two genotypes against the target quantitative trait.
All the genetic markers except for G10 have variances equal to zero or less than a
sufficiently small number. Their effects of the genotypes (uj) also have similar
characteristics. Thus, the genetic markers with such small effects can be excluded
from the linear model, matching the expectation initially made for the dataset in
Fig. 3.1.
If we exclude the genetic markers with small variance (σ 2j ) and coefficients (uj)
other than the 10th genetic marker (σ 21 ¼ 0.04) from the model, we have:
0 1
β0
@ A u11
y ¼ X β1 þ Z10 þ e: ð3:97Þ
u12
β2
Using the values obtained by the Gibbs sampling (Fig. 3.2), we have:
0 1
0:59
@ A 0:19
y ¼ X 0:25 þ Z10 þ e: ð3:98Þ
0:31
0:03
If a sample has the genotype encoded as 1 (gray color in Fig. 3.1) in the design
matrix at the 10th genetic marker, the value by the model is 0.59 + 0.19 ¼ 0.78. On
the other hand, if a sample has the genotype encoded as 0 in the design matrix at the
10th genetic marker, the value by the model is 0.590.31 ¼ 0.28. The contributions
of the environmental factors, E1 and E2, are, respectively, 0.25 and 0.03 and
cause a difference in the trait value of 0.25 + 0.03 ¼ 0.28, meaning that the samples
in E1 have a gain of 0.28 relative to E2. Using the model obtained, we can predict
the trait value of a sample.
3.8 Summary and Conclusions
A typical framework of GS requires prediction of the trait values of the samples to

determine the candidates for the selected individuals. Finding the genetic markers
associated with a target trait might not be sufficient for GS, even if the genetic
markers are distributed in a genome-wide manner. Therefore, deriving the predic-

tion model of the target trait is an essential part of GS, since it enables the selection
to be based on the predicted trait values. Once such a prediction model is
established, the samples with the identical genetic characteristics can be evaluated
even if they do not have the trait values observed. When, however, the number of
genetic markers is excessive relative to the number of the samples, or the number of
the samples is not sufficient for the number of the genetic markers, several assump-
tions for the estimation of the model parameters—e.g., the genetic markers are
independent of each other—break and the prediction model will be distorted.
Linkage disequilibrium (LD) actually appears among the genetic markers in a
dataset with a small number of samples. Although it is not clear that such depen-
dence among the genetic markers reflects the true biological or genetic mecha-
nisms, the number of the genetic markers sharing the same or a similar genotype
pattern increases at least in those small datasets, detracting from the prediction
ability. The problems related to overfitting must also be taken into account to assure
the generalization ability. The accuracy in the prediction of the hidden trait values
of the samples that are independent from the model construction can be evaluated
by methods such as the cross-validation test and the bootstrapping test. Although it
will be necessary to carry out further inspection of the prediction model for
practical use, the strategy based on the linear mixed models and the Bayesian
estimation will be useful as a fundamental step in the prediction of trait values of
samples based on their genetic and environmental factors.
References
Box GEP, Muller ME (1958) A note on the generation of random normal deviates. Ann Math
Statist 29(2):610–611
wide dense marker maps. Genetics 157(4):1819–1829
Tanizaki H (2008) A simple gamma random number generator for arbitrary shape parameters.
Econ Bull 3(7):1–10
Wang CS, Rutledge JJ, Gianola D (1993) Marginal inferences about variance components in a
mixed linear model using Gibbs sampling. Genet Sel Evol 25(1):41–62
Chapter 4
Bayesian Genomic-Enabled Prediction Models
for Ordinal and Count Data
Osval A. Montesinos-López, Abelardo Montesinos-López, and José Crossa
4.1 Introduction
Animal and plant breeding have been revolutionized by genomic-enabled predic-

tion models. This tool is powerful for predicting the genomic merit of animals and
plants based on high-density single nucleotide polymorphism (SNP) marker panels,
and it has been implemented for genomic prediction for the predisposition to some
diseases in human health (Yang and Tempelman 2012). However, most existing
genomic-enabled prediction models assume normality (in the phenotype and error),
linearity in the model parameters, and a constant variance. To translate these
models for a non-Gaussian context is a complex task because integrating over the
random effects is intractable (McCulloch and Searle 2001). For this reason,
researchers normally approach non-Gaussian phenotypes in three ways: (a) they
assume normality in the phenotypes, (b) they approximate to normality the
non-normal response transforming the phenotype, or (c) they model the appropriate
distribution of the phenotype using generalized linear mixed models (GLMMs)
(Stroup 2015).
The first approach is justified for large sample sizes by the central limit theorem.
However, empirical and simulation studies have shown that this first approach
produces highly biased results for small and moderate sample sizes (Stroup 2012,
O.A. Montesinos-López
Facultad de Telemática, Universidad de Colima, Colima 28040, Colima, Mexico
A. Montesinos-López
Departamento de Matemáticas, Centro Universitario de Ciencias Exactas e Ingenierı́as
(CUCEI), Universidad de Guadalajara, Jalisco, Guadalajara 44430, Mexico
J. Crossa (*)
Biometric and Statistics Unit (BSU), International Maize and Wheat Improvement Center
(CIMMYT), Apdo. Postal 6-641, 06600, D.F., México
e-mail: j.crossa@cgiar.org

DOI 10.1007/978-3-319-63170-7_4
56 O.A. Montesinos-López et al.
2015). Transformations originally were proposed for variance stabilization of

non-normal data to obtain a homogeneous variance (Bartlett 1947); this approach
is still popular in many agricultural disciplines. The transformed phenotypes are
assumed normally distributed variables and are implemented with the traditional
linear model. However, often these remedial measures produce a great loss of
accuracy and power (Stroup 2015), mostly in small sample sizes.
GLMMs unify models characterized as being linear on the systematic compo-
nent (model predictors). For this reason, they are appropriate for normal and
non-normal data with heterogeneous variance and even correlated observations
(Nelder and Wedderburn 1972). GLMMs are very popular in many areas (finance,
healthcare, biostatistics, etc.) but, to date, still underutilized in the agricultural
research community. Empirical and simulation studies on small sample investiga-
tions show that GLMMs produce more accuracy and power than approaches (a) and
(b) previously described. Also, for implementing GLMMs there are textbooks and
software available, although implementation of approaches (a) and (b) are the
dominant approaches in agricultural research (Stroup 2015). The use of GLMMs
in genomic-enabled prediction is new because their implementation is not straight-
forward given that the number of observations usually is smaller than the number of
covariates. In addition, the joint involvement of biological processes and pathways
complex dependence structures are observed among markers and lines.
In the pre-genomic era, the use of models for non-normal data is not new. Wright
(1934) developed the threshold concept to map a normally distributed underlying
variable to the observed categorical phenotypes, and the ordinal categorical phe-
notype is assumed to be the visible expression of an underlying continuous variable
(de Maturana et al. 2009). Gianola (1980, 1982) and Gianola and Foulley (1983)
proposed a probit (threshold) model for ordinal categorical traits in animal breeding
and Gonzalez-Recio, Forni (González-Recio and Forni 2011) and Villanueva et al.
(2011) for binary trials. Authors Wang et al. (2013) and Montesinos-López et al.
(2015a) extended to threshold model for more than two ordinal categories to deal
with p n in the genomic era. Also, de los Campos and Perez-Rodriguez (2013)
developed the BGLR package for genomic-enabled prediction for normal, binary,
ordinal, and censored data. A log transformation is often used for counts to satisfy
normality rather than being modeled on the basis of a count distribution. This
transformation for count data is inefficient when there are zeros as observations,
because with only one observation with zero, the entire data set needs to be shifted
by adding an arbitrary value (usually 1) before transformation. Also, many times
this transformation performs poorly, except when dispersion is small and mean
counts are large (O’Hara and Kotze 2010). Next we present a review of the existing
methods for genome-enabled prediction models for ordinal and count data that give
a better idea of the need to develop this type of models.
Kizilkaya et al. (2014), in their paper titled “Reduction in accuracy of genomic
prediction for ordered categorical data compared to continuous observations,”
pointed out that methods used to analyze continuously distributed traits are not
optimal for analyzing categorical traits; therefore, it is important to develop appro-
priate methods for categorical ordinal data. Many low heritability traits have
4 Bayesian Genomic-Enabled Prediction Models for Ordinal and Count Data 57
ordered categorical scores, such as susceptibility or resistance to a disease and

reproductive traits such as calving difficulty (Kizilkaya et al. 2014). The goal of this
study was to quantify reductions in accuracy for ordinal categorical traits relative to
continuous traits (Kizilkaya et al. 2014).
For the above reasons, Kizilkaya et al. (2011) used a BayesC threshold model to
analyze the ordinal categorical trait infectious bovine keratoconjunctivitis in Angus
beef cattle. The same model was used for the genome-wide association analysis of
pregnancy in Brangus heifers and first service conception (González-Recio and
Forni 2011) and for insect bite hypersensitivity (Schurink et al. 2012).
Kizilkaya et al. (2014) show that genome-wide analysis of ordinal categorical data
produced substantially lower accuracy of genomic expected breeding values (GEBV)
than the analysis of a continuous phenotype. Kizilkaya et al. (2014) also found that a
2.25 larger training population size for ordinal categorical phenotypes analyzed using
a threshold model is required to achieve an accuracy equal to or greater than that for
continuous phenotypes for a training population size of 1000 animals. However,
using a linear model (assuming normality), a more than 2.25-fold increase in the size
of the training population would be required to achieve the same accuracy as a
continuous trait with 1000 observations for analyzing an ordinal categorical pheno-
type. They also found that GEBV accuracy increased significantly when the training
population size and heritability increased for the threshold model and for all number
of categories in the ordinal categorical data (Kizilkaya et al. 2014).
Kizilkaya et al. (2014) also concluded that when analyzing categorical data, the
threshold model had higher accuracies than the linear model (which assumes
normality in the phenotype). The research of Varona et al. (1999) also reached
similar conclusions when comparing linear and threshold models in conventional
pedigree-based evaluations (EBV) using simulated data sets for calving difficulty.
A study of Ramirez-Valverde et al. (2001) also supports this finding; they compared
EBV accuracy of threshold animal, threshold sire-maternal grandsire, linear animal,
and linear sire-maternal grandsire models for calving difficulty in beef cattle and
determined that EBV accuracy of the threshold model was 10% higher than EBV
accuracy of the linear model for animal and sire-maternal grandsire models. Casella
et al. (2007) analyzed litter size using linear and threshold models and found better
goodness of fit and predictive ability for EBV in a threshold model than in a linear
model (Kizilkaya et al. 2014). These results are in agreement with those reported by
Villanueva et al. (2011), who developed a version of the BayesB method for
dichotomous traits and concluded that the threshold BayesB method improves
prediction accuracy when dealing with disease-resistant dichotomous phenotypes,
compared with accuracies obtained with the linear model. The threshold model
showed an increase in accuracy of up to 16%, as well as significant advantages
when heritability and disease prevalence were low and individuals were genotyped
but not measured (testing set).
Kizilkaya et al. (2014) concluded that bias in predictions is reduced in the
threshold model when heritability and training population size increase. Although
this bias is considerable, it is worse when the data are categorical ordinal but
analyzed as if they were continuous using a linear model. Kizilkaya et al. (2014)
also point out that linear model analyses perform as well as threshold model
analyses when the number of categories is large. However, when training
populations are small, the accuracies of GEBV for ordinal categorical phenotypes
analyzed by the threshold model are higher than those analyzed with a linear model
applied to the ordinal data.
On the other hand, Wang et al. (2013) showed that Bayesian threshold methods
(BayesTA, BayesTB, and BayesTCπ) performed better than the corresponding
normal Bayesian methods (BayesA, BayesB, and BayesCπ) in all cases (with
20, 50, 200, and 500 QTL), except in the case of 20 QTL, where BayesB, BayesCπ,
BayesTB, and BayesTCπ gave almost the same accuracies. Wang et al. (2013) also
found that BayesTB, BayesTCπ, BayesB, and BayesCπ were sensitive to the
number of QTL, and their accuracies decreased rapidly when the number of
simulated QTL increased from 20 to 200. In contrast, BayesTA and BayesA
accuracies did not change (were not sensitive) to the number of simulated QTL.
Wang et al. (2013) also found that when the incidence of the binary trait
decreased from 50% to 5%, the accuracies of GEBV decreased consistently. But
the three Bayesian threshold methods (BayesTA, BayesTB, and BayesTCπ)
performed better than the corresponding normal Bayesian methods in all cases.
BayesTB and BayesTCπ produced similar accuracies, and their advantage over
BayesB and BayesCπ increased as incidence decreased. Wang et al. (2013) also
found that as the number of phenotypic categories increased, the accuracies of
GEBV for all the Bayesian methods increased, but the superiority of the three
BayesT methods over the corresponding normal Bayesian methods decreased as the
number of categories increased, and with eight or more categories, the three BayesT
methods completely lost their advantage. BayesA was the most sensitive of all
methods to the number of categories, while BayesTA was not sensitive to the
number of categories.
Wang et al. (2013) found that the accuracies in generation 2 improved by 30.4%,
2.4%, and 5.7% for BayesTA, BayesTB, and BayesTCπ, respectively, when the
number of categories ¼ 2, incidence ¼ 0.3, number of QTL ¼ 50, and heritabil-
ity ¼ 0.3. They also concluded that the performance of the methods (threshold and
normal) significantly is affected by the genetic architecture underlying the traits
since the accuracies of all methods declined with the decrease of the heritability
when increasing the number of QTL.
The threshold model above mentioned were developed and applied in the
context of animal breeding. However, the study by Montesinos-López et al.
(2015a) titled “Threshold models for genome-enabled prediction of ordinal cate-
gorical traits in plant breeding,” was conducted, in the context of plant breeding.
Montesinos-López et al. (2015a) extended the so-called genomic best linear unbi-
ased predictor (GBLUP) model for Gaussian phenotypes to ordinal data with probit
link (TGBLUP). The main contributions of Montesinos-López et al. (2015a) are
summarized as:
(a) Real data were used, not simulated data as in Wang et al. (2013) and Kizilkaya
et al. (2014).
(b) They take into account genotype environment interaction (G E)

interaction.
(c) They provide a very clear description of the threshold model and provides R
code for its implementation.
(d) They provide an alternative metric (Brier score) for assessing prediction accu-
racy for categorical ordinal outcomes.
(e) They take into account epistatic additive additive terms, even though this did
not help much to increase prediction accuracy.
Montesinos-López et al. (2015a) found that models that take into account G E
have higher prediction accuracy than those that ignore the G E term. Relative to
models based on main effects only, models that include G E gave gains in
prediction accuracy between 9% and 14%.
Due to their connection to odds ratios and since it provides regression coeffi-
cients that are more interpretable, the ordinal logistic regression model is often
preferred over the ordinal probit model in statistical applications (Zucknick and
Richardson 2014). However, only the Bayesian probit ordinal regression (BPOR)
model is frequently implemented in genomic-enabled prediction (when p n),
given that Bayesian methods that introduce sparseness through additional priors on
the model size are very well suited to this problem. Due to the lack of a Bayesian
logistic ordinal regression (BLOR) model analogous to the BPOR model that uses a
data augmentation approach, Montesinos-López et al. (2015b) proposed the BLOR
with logit link, without taking into account G E interaction. This BLOR model
was developed using the Pólya-Gamma data augmentation approach that produces
a Gibbs sampler with similar full conditional distributions as the BPOR model, with
the advantage that the BPOR model is a particular case of the BLOR model. The
authors evaluated the proposed BLOR model using three sets of data. Results from
Montesinos-López et al. (2015b) indicate that BLOR model is an alternative for
analyzing ordinal data in the context of genomic-enabled prediction with the probit
or logit link.
For count data, only two models for genomic-enabled prediction were found.
The first one was proposed by Montesinos-López et al. (2015c) titled “Genomic
prediction models for count data,” extended the GBLUP to count data (CGGLUP),
and allows modeling count data without assuming that the data are normally
approximated and without using transformation, which many times produces esti-
mations and predictions outside of nonnegativity, which makes no sense for count
data. However, this model does not take into account G E interaction. For this
reason, Montesinos-López et al. (2016) extended this model to incorporate G E
interaction. They found that the Bayesian negative binomial regression (BNBR)
model with G E improved prediction accuracy compared to the normal model
and to a normal model that uses log-transformed responses.
For both models (BLOR and BNBR), there are Bayesian implementations that
estimate the posterior distribution of the required parameters via the Markov chain
Monte Carlo (MCMC) algorithm with Gibbs sampling; however, full conditionals
for the above models do not have analytic solutions. Therefore, this approach is not
useful for large data sets that are commonly used in genomic selection.
For the above reasons, in this chapter, we provide an extension of the BLOR
taking into account G E interaction for ordinal categorical phenotypes; we also
provide details of the derivation and implementation of the BNBR model for count
data proposed by Montesinos-Lopez et al. (Montesinos-López et al. 2016) that
include the G E term. The full conditionals of the parameters required in both
models were obtained analytically using the Pólya-Gamma augmentation approach
(Polson et al. 2013) that allows the implementation of an efficient Gibbs sampler for
the BLOR and BNBR models with G E. These models could be very useful for
genomic-enabled prediction in plant breeding because they take into account G E
interaction and are powerful enough to deal with large numbers of covariates and
small numbers of observations. We illustrate our proposed method with simulation
and a real data set.
4.2 Materials and Methods
4.2.1 Data Sets
4.2.1.1 Gray Leaf Spot and Septoria Data Sets
Gray leaf spot (GLS), caused by Cercospora zeae-maydis, is a foliar disease of

global importance in maize production. The disease was evaluated using an ordinal
scale [1 (no disease), 2 (low infection), 3 (moderate infection), 4 (high infection),
5 (complete infection)] in three environments (Mexico, Harare, and Colombia). Of
the 278 maize lines evaluated, only 240 were the same in the three environments.
For this reason, we used only the 240 lines to illustrate our methods with real data.
The use of a Poisson random variable for analyzing ordered categorical responses is
not new; for example, Vazquez et al. (2009) compared Poisson and threshold
models for genetic analysis of clinical mastitis in the US Holsteins. These data
are part of a data set that was previously analyzed (Crossa et al. 2011; González-
Camacho et al. 2012) under the assumption of normality and using a threshold
model for ordinal data (Montesinos-López et al. 2015a; Montesinos-López et al.
2015b). Genotypes of all 240 lines used were obtained using the 55 k single
nucleotide polymorphism (SNP) Illumina platform. SNPs with >10% missing
values or a minor allele frequency of 0.05 were excluded from the data. After
line-specific quality control (applying the same quality control to each line sepa-
rately), the maize data still contained 46,347 SNPs, which were used in the analysis.
4.2.1.2 Fusarium Head Blight Data
From a total of 297 spring wheat lines from CIMMYT evaluated for resistance to
Fusarium head blight (FHB), 182 were used for implementing the models for count
data because only for these lines had complete marker information. The
phenotyping was measured in three environments (El Batan 2012, El Batan 2014,
and Ecuador 2014). In each environment the genotypes were arranged in a ran-
domized complete block design. The response variable FHB severity data were
collected shortly before maturity by counting symptomatic spikelets on ten ran-
domly selected spikes in each plot. DNA samples were genotyped using an Illumina
9 K SNP chip with 8632 SNPs (Cavanagh et al. 2013). After filtering the markers
for 0.05 minor allele frequency (MAF) and deleting markers with more than 10% of
no calls, the final set of SNPs was 1635 SNPs (Montesinos-López et al. 2016). This
data set was only used for the models for count data.
4.2.2 Statistical Models
We use yijt to represent the response for the tth replication of the jth line in the ith
environment with i ¼ 1 , . . . , I ; j ¼ 1 , 2 , . . . , J , t ¼ 1 , 2 , . . . , nij, and we propose
the following linear predictor that takes into account G E:
ηij ¼ Ei þ gj þ gEij ð4:1Þ
where Ei represents the environment i and is assumed fixed, gj is the marker effect
of genotype j, gEij is the interaction between genotypes and environments, I ¼ 3
(Colombia, Zimbabwe, and Mexico), J ¼ 240 (i.e., the number of lines under
study), and nij represent the number of replicates of each Xline in each environment.
J
The number of observations in environment i is ni ¼ n , while the total
j¼1 ij
XI
number of observations is n ¼ n : Rewriting the linear predictor (Eq. 4.1) as
i¼1 i
ηij ¼ xTi β þ b1j þ b2ij ð4:2Þ
with xTi ¼ ½xi1 , xi2 , xi3 , where xi1 , xi2, and xi3 are indicator variables that take the
value of 1 if the observed environment i is 1, 2, or 3, respectively, and 0 other-
wise, βT ¼ [β1, β2,, β3,] because three is the number of environments under study,
xTi β ¼ Ei , b1j ¼ gj and b2ij ¼ gEij. Three models are proposed using the linear
predictor given in Eqs. (4.1) and (4.2).
Model BLOR In this model, the linear predictor is ηij(c) ¼ γ c ηij, where ηij(c) denotes
the cth link (c ¼ 1 , 2 , . . . , C 1) for the fixed and random effects combination, γ c is the
threshold (intercept) for the cth link, and ηij is exactly as defined in Eq.(4.2). Distributions:
yijt(1) , yijt(2) , . . . ,yijt(C)|ηij~Multinomial(1 , π ij(1) , π ij(2) , . . . ,π ij(C)). b1 ¼ ðb11 , . . . , b1J ÞT
T T
Nð0, G1 σ 2b1 Þ, b2i ¼ (b2i1, . . . , b2iJ)T, b2 ¼ ðb21 T
, . . . , b2I Þ Nð0, G2 σ 2b Þ. Link func-
2
π ijðcÞ
tion: cumulative logit { ηijðcÞ ¼ log 1π ijðcÞ , c ¼ 1 , 2 , . . . , C 1)}, since there are
C categories, a total of C 1 link functions are required to fully specify the model. G1
and G2 were assumed known, with G1 computed from marker W data (for m ¼ 1, . . . ,
q markers) as G1 ¼ WW
T
q ; this matrix is called the genomic relationship matrix (GRM)
(VanRaden 2008). The G1 matrix is a covariance matrix that contains the similarity
between individuals based on marker information, rather than on expected similarity
based onpedigree,
N that can help to improveprediction
N accuracy. While G2 is computed as
G2 ¼ II G1 of order IJxIJ and denotes the Kronecker product, II means that we
assume independence between environments (Montesinos-López et al. 2016).
Model BNBR Linear predictor as given in Eq. (4.1). Distributions: yijt|ηij~NB(μij , r),
NB stands for negative binomial distribution with r being the dispersion parameter
(shape parameter), μij ¼ E(yijt|ηij) ¼ exp(ηij). Note that the BNBR has expected value
rπ rπ ij μ2ij
μij ¼ 1πij and variance ¼ μ þ , with the variance greater than the
ð ij Þ 1 π ij
2 ij
r
mean (Montesinos-López et al. 2016).
Model Pois Everything is the same as in model BNBR, except that yijt|ηij~Poisson
(μij). Since according
to Zhou
et
al. (2012) and Teerapabolarn and Jaioun (2014),
the lim NB μij ; r ¼ Pois μij , model Pois was implemented using the same
r!1
method as model BNBR, but fixing r to a large value, depending on the mean
count. We used r ¼ 1000, which is reasonable when the mean count is less than
50 (see Fig. 4.1). However, for mean counts between 50 and 200, we suggest using
r ¼ 5000, and for counts larger than 200, we suggest a value of r ¼ 10000 or larger.
Figure 4.1 supports these suggestions where we plot the mean and variance of
200
Fig. 4.1 Plot of the mean

count versus the variance of r= 1000
NB distribution as a function r= 5000
of the scale parameter (r). r= 10000
μ= σ2
150
Good approximations are

obtained when the mean and
σ2 (Variance)
variance are very similar; in

the plot, they should follow
100
the diagonal that plots μ ¼ σ 2

(Extracted from
Montesinos-López et al.
2016)
50
0
0 50 100 150 200

μ (Mean)
model BNBR as a function of the scale parameter r, with three values of r (1000,
5000, 10,000). Acceptable approximations to the model Pois with the model
BNBR occur when the mean and variance are very similar. For this reason, good
approximations are those that follow the diagonal in Fig. 4.1, where μ ¼ σ 2. The
mean count and variances are very similar for mean counts of less than 50 with
r ¼ 1000; however, when the mean count is larger than 50 and less than 200, we
should use r ¼ 5000, and for counts greater than 200, we suggest using a value of
r ¼ 10,000 or larger. In our applications with simulated and real data, the mean
count is less than 50; for this reason, we used a value of r ¼ 1000. Next, we provide
details of the derivation of the full conditional distribution for each model
(Montesinos-López et al. 2016).
4.2.2.1 Bayesian Logistic Ordinal Regression (BLOR)

h iT T T
Let yij ¼ yij1 ; . . . ; yijnij , yi ¼ yi1T ; . . . ; yiJT , and y ¼ y1T ; . . . ; yIT ; in this
model, the response variable yijt represents an assignment into one of C mutually
exclusive and exhaustive categories that follow an order. Therefore, the ordinal
logistic regression model can be written in terms of a latent response variable lijt as
lijt ¼ xTi β þ b1j þ b2ij þεijt ð4:3Þ
where lijt are called “liabilities,” εijt ~ L(0, 1), where L(.) denotes the logistic distri-
bution, and the remaining terms are as defined in Eq. (4.2). Since lijt are
unobservable and can be measured indirectly by an observable ordinal variable
yijt, then yijt can be defined by
8
>
> 1 if 1 < lijt < γ 1 ,
<
2 if γ 1 < lijt < γ 2 ,
yijt ¼
>
> ⋮
:
C if γ C1 < lijt < 1
This means that lijt is divided by thresholds into C intervals, corresponding to

C ordered categories. The first threshold, γ 1, defines the upper bound of the interval
corresponding to observed outcome 1. Similarly, threshold γ C 1 defines the lower
bound of the interval corresponding to observed outcome C. Threshold γ c defines the
boundary between the interval corresponding to observed outcomes c 1 and c for
(c¼1,2,.., C -1). Threshold parameters are γ T ¼ (γ min < γ 1 < < γ C 1 < γ max) with
γ min ¼ 1 and γ max ¼ 1 .
Because that the error term εijt of the latent response lijt is distributed as L(0, 1),
the cumulative response probability for the c category of the ordinal outcome yijt is
Pðyijt cjβ, b1 , b2 Þ ¼ π ijðcÞ ¼ Pðlijt γ c jβ, b1 , b2 Þ ¼ PðxTi β þ b1j þ b2ij þ εijt γ c Þ

¼ Pðεijt γ c xTi β b1j b2ij Þ, for c ¼ 1, 2, . . . , C 1:
expðγ c xTi β b1j b2ij Þ
¼ ð4:4Þ
1 þ expðγ c xTi β b1j b2ij Þ
Similarly, Eq. (4.4) can be written as a cumulative logit model:

π ijðcÞ
log Þ ¼ γ c xTi β b1j b2ij , for c ¼ 1, 2, . . . , C 1:
1 π ijðcÞ
Using the inverse link for this model, P(yijt ¼ c|β, b1, b2) ¼ π ij(c) can be calculated
as follows:
π ijðcÞ ¼ Pðγ c1 < lijt < γ c Þ

expðγ c xTi β b1j b2ij Þ expðγ c1 xTi β b1j b2ij Þ
¼ :
1 þ expðγ c xTi β b1j b2ij Þ 1 þ expðγ c1 xTi β b1j b2ij Þ

Since we have latent variables lijt distributed as L xiT β þ b1j þ b2ij ; 1 and we
observe yijt ¼ c if, and only if, γ c 1 < lijt < γ c, then the joint posterior density of the
parameter vector and latent variable becomes
f ðβ, γ, b1 , b2 , σ 2β , σ 2b1 , σ 2b2 ljyÞ / f ðyjl, γÞf ðθT Þ

!
where f ðθT Þ ¼ f ðljβ, b1 , b2 Þf ðγÞf βjσ 2β f ðb1 jσ 2b1 Þf ðb2 jσ 2b2 Þf ðσ 2β Þf ðσ 2b1 Þf ðσ 2b2 Þ: Then
2
! χ (νh, Sh) prior for σ bh for
2
assuming a scaled independent inverse chi-square
h ¼ 1 , 2, a normal prior distribution for f βjσ 2β Nðβ0 , Σ0 σ 2β Þ, a normal prior
distribution for f ðb1 jσ 2b1 Þ N J ð0, G1 σ 2b1 Þ, a normal prior distribution for
f ðb2 jσ 2b2 Þ N IJ ð0, G2 σ 2b2 Þ, and also a χ 2(νβ, Sβ) prior was given for σ 2β (Gianola
2013). Following Sorensen et al. (1995), a prior for the C 1 unknown thresholds
has been given as order statistics from U(γ min, γ max) distribution,
C1
1
Pðγ Þ ¼ ðC 1Þ! I ðγ 2 TÞ
γ max γ min
where T ¼ {(γ 1, . . . , γ max)|γ min < γ 1 < < γ C 1 < γ max}.

The full conditional posterior distributions are provided below and in Appendix
A are all details of these derivations.
4.2.2.2 Liabilities and ωijt
The fully conditional posterior distribution of liability lijt is a truncated normal

distribution and its density is
f ðlijt jELSEÞ
pffiffiffiffiffiffiffi
ωijt ðlijt xTi β b1j b2ij Þ
ϕ
¼ pffiffiffiffiffiffiffi
Φð ωijt ðγ c xTi β b1j b2ij ÞÞ Φ ωijt ðγ c1 xTi β b1j b2ij Þ
ð4:5Þ
For simplicity, ELSE is the data and the parameters, except for the one in
question. ϕ and Φ are the density and distribution function of a standard normal
random variable, and the full conditional posterior distribution of ωijt is
f ðωijt jELSEÞ PGð2, lijt þ xTi β þ b1j þ b2ij Þ ð4:6Þ
where PG stands for the Pólya-Gamma distribution.
4.2.2.3 Regression Coefficients (β)
The full conditional posterior of β is

f βELSE N p ~ ~0
β0; Σ ð4:7Þ
P2
where Σ ~ 0 ¼ ðΣ1 σ 2 þ XT Dω XÞ1 , and ~ β0 ¼ Σ ~ 0 ðΣ1 σ 2 β0 XT Dω Zh bh þ
0 β 0 β
T T h h¼1
iT
XT Dω lÞ. With l ¼ l1T ; . . . ; lIT , li ¼ li1T ; . . . ; liJT , lij ¼ lij1 ; . . . ; lijnij ,
h O iT T T
Xij ¼ 1nTij xi , Xi ¼ Xi1T ; . . . ; XiJT , X ¼ X1T ; . . . ; XIT ,

Dωij ¼ diag ωij1 ; . . . ; ωijnij , Dωi ¼ diag(Dωi1, . . . , DωiJ), Dω ¼ diag(Dω1, . . . , DωI),
T T
b1 ¼ [b11, . . . , b1J]T, b2i ¼
2 3 [b2i1, . . . , b2iJ] ,
T
b2 ¼ b21 ; . . . ; b2IT , Z1i ¼
1n1i1 0 0
6 0 1n1i2 0 7
6 7, Z1 ¼ Z T ; . . . ; Z T T , and Z2 ¼ Z1∗ ~ X, where ∗~indi-
4 ⋮ ⋮ ⋱ ⋮ 5 11 1I
0 0 1n1iJ
cates the horizontal Kronecker product between Z1 and X. The horizontal Kronecker
product performs a Kronecker product of Z1 and X and creates a new matrix by stacking
these row vectors into a matrix. Z1 and X must have the same number of rows, which is
also the same number of rows in the resulting matrix. The number of columns in the
resulting matrix is equal to the product of the number of columns in Z1 and X. It is
important to point out that if we use a prior for β / Constant (improper uniform
distribution), then in Σ ~ 0 and ~
β 0 we need to make 0 the term Σ1 2
0 σβ .
4.2.2.4 Polygenic Effects (bh)
The full conditional distribution of bh with h ¼ 1 , 2, is given as

f ðbh jELSEÞ N bh ¼ Fh ðZhT Dω l ZhT Dω ηh Þ, Fh

1 T 1
¼ ðσ 2
bh Gh þ Zh Dω Zh Þ
T
ð4:8Þ
with η1 ¼ Xβ + Z2b2 and η2 ¼ Xβ + Z1b1.
4.2.2.5 Variance of Polygenic Effects ðσ 2bh Þ
Next, the full conditional posterior of σ 2bh is

f ðσ 2bh jELSEÞ χ 2 νh ¼ νh þ nbh , Sb ¼ ðbhT G1
h bh þ νh Sh Þ=νb þ nbh Þ ð4:9Þ
with nb1 ¼ J and nb2 ¼ IJ.
4.2.2.6 Threshold Effects (γ c)
The density of the full conditional posterior distribution of the cth threshold, γ c, is
f ðγ c jELSEÞ
1
¼
minfminðlijt jyijt ¼ c þ 1Þ, γ cþ1 , γ max g maxfmaxðlijt jyijt ¼ cÞ, γ c1 , γ min g
ð4:10Þ
4.2.2.7 Variance of Regression Coefficients
The full conditional posterior of σ 2β is

h i
f σ 2β jELSE χ 2 ~ν β ¼ νβ þ p; ~
S β ¼ ðβ β0 ÞT Σ 1
0 ðβ β0 Þ þ νβ Sβ =νβ þ p
ð4:11Þ
4.2.2.8 The Gibbs Sampler for model BLOR
The Gibbs sampler is implemented by sampling repeatedly from the following

loop:
Step 1. Sample liabilities (lijt) from the truncated normal distribution in (4.5).
Step 2. Sample ωijt values from the Pólya-Gamma distribution in (4.6).
Step 3. Sample the regression coefficients (β) from the normal distribution in (4.7).
Step 4. Sample the polygenic effects (bh) for h ¼ 1 , 2, from the normal distribution
in (4.8).
Step 5. Sample the variance effect (σ 2bh Þ for h ¼ 1 , 2,from the scaled inverted χ 2
distribution in (4.9).
Step 6. Sample the thresholds (γ c) from the uniform distribution in (4.10).
Step 7. Sample the variance of regression coefficients (σ 2β Þ from the scaled inverted
χ 2 distribution in (4.11).
Step 8. Return to step 1 or terminate if chain length is adequate to meet convergence
diagnostics.
Ignoring the polygenic effects (bh), the Gibbs sampler given above can be used
only by deleting steps 4 and 5. If all marker effects are considered fixed effects and
included in the design matrix, X, with a prior β N p ð0, Ip σ 2β Þ for the beta regression
coefficients, we end up with a threshold Bayesian ridge regression. This is the ridge
estimator of Hoerl and Kennard (1970) for ordinal categorical data, since the
1
posterior expectation of β is equal to Eðβj ELSEÞ ¼ XT Dω X þ Ip σ 2 β XT Dω l
with pseudo-response l. Also, note that setting each ωijt ¼ 1, the Gibbs sampler
given above for the BLOR with the logistic link is reduced to the Gibbs sampler
with the probit for the BPOR link proposed by Albert and Chib (1993). Therefore,
our proposed BLOR model is more general and includes the Gibbs sampler for the
BPOR model as a particular case as implemented in BGLR package.
4.2.3 Bayesian Mixed Negative Binomial Regression
Note that under the model BNBR, because μij ¼ E(yijt|ηij) ¼ exp(ηij), conditionally
on b1j and b2ij, the probability that the random variable Yijt takes the value yijt is
equal to
!r !yijt
yijt þ r 1 μij μij
Pr Y ijt ¼ yijt ηij ¼ 1 for yijt
yijt r þ μij r þ μij
¼ 0, 1, 2, . . .
h iyijt
Γ yijt þ r exp η∗
ij
¼ h i , yijt ¼ 0, 1, 2, . . . ð4:12Þ
yijt !Γ ðr Þ 1 þ exp η∗ yijt þr
ij
μij
We arrive at Eq. (4.12) because we make ¼rμij
rþμij

μij=r exp ηij expðlogðr ÞÞ exp ηij logðr Þ exp η∗
ij
r rþμij ¼ ¼ ¼ ¼ ,
1þμij=r 1þexp ηij expðlogðr ÞÞ 1þexp ηij logðr Þ 1þexp η∗
ij
T ∗ ∗
∗ ∗ ∗
with η∗ ∗
ij ¼xi β þb1j þb2ij , β ¼ β 1 ;β 2 ;β3 and β i ¼β i logðr Þ. Therefore, in
Eq. (4.12) we have the connection between the probability distribution of the response
(Yijt) induced by the assumed relation between the linear predictor (ηij) and the expected
value of Yijt (μij) under model BNBR. Then we can rewrite the Pr(Yijt¼yijt|ηij) given in
Eq. (4.12) as
2 2 3
Γ yijt þ r y r ð 1 ω η∗
6 ijt ij 7
2yijt r exp η∗
ijt
exp 4 5f ωijt , yijt þ r; 0 dωijt
yijt !Γ ðr Þ 2 ij
0 2
ð4:13Þ
Expression (4.13) was obtained ð 1using the following equality given by Polson
ðeψ Þa ω ψ2
b κψ ijt2
et al. (2013): b
¼2 e e f ωijt ; b; 0 dωijt , where κ ¼ a b/2 and
ð1 þ eψ Þ 0
f(., b, 0) denotes the density of the Pólya-Gamma distribution (ωijt) with parameters
b and c ¼ 0 (PG(b, c ¼ 0)) (see Definition 1 in Polson et al. 2013). From here,
conditioning on ωijt ~PG(yijt + r, c ¼ 0), we have that

Γ yijt þ r y r 2

Pr Y ijt ¼ yijt ηij ; ωijt ¼ 2 yijt r
exp
ijt
ηij exp ωijt η∗
∗
=2
yijt !Γ ðr Þ 2 ij
ð4:14Þ
To be able to get the full conditional distributions, here we provide the prior
distributions, f(θ), for all the unknown model parameters β∗, σ 2β∗ , b1, σ 2b1 , b2, σ 2b2 ,
and r. We assume the following prior between the parameters, that is,
!
∗
f ðθÞ ¼ f β jσ 2β∗ f ðσ 2β∗ Þf ðb1 jσ 2b1 Þf ðσ 2b1 Þf ðb2 jσ 22 Þf ðσ 2b2 Þf ðrÞ
We assign conditionally conjugate but weakly informative prior distributions to

the parameters because we have no prior information. The prior specification in
terms of β∗ instead of β is for convenience. We adopt proper priors with known
hyper-parameters whose values we specify in modelimplementation to guarantee

X
proper posteriors. We assume that f β∗ σ 2β∗ N I β0 ; σ
0 β
2
, f σ 2β∗ χ 2

νβ ; Sβ where χ 2 νβ ; Sβ denote a scaled inverse chi-square distribution with
shape νβ and scale S!
β parameters, f ðb1 jσ 2b1 Þ N J ð0, G1 σ 2b1 Þ,

f σ 2b1 χ 2 ðνb1 ; Sb1 Þ, f b2 jσ 2b2 N IJ ð0, G2 σ 2b2 Þ, f σ 2b2 χ 2 ðνb2 ; Sb2 Þ, and f
(r) ~ G(a0, 1/b0). Next we combine (4.14) using all data with priors to get the full
conditional distribution for parameters β∗, σ 2β∗ , b1, σ 2b1 , b2, σ 2b2 , and r.
4.2.4 Full Conditional Distributions
The full conditional distribution of β∗ is given as

fðβ∗ jELSEÞ N ~ ~0
β0; Σ ð4:15Þ
1 X
2
~ 0 ¼ Σ1 σ 2 þ XT Dω X
where Σ 0 β , ~ ~ 0 Σ1 σ 2 β0 XT Dω
β0 ¼ Σ 0 β Zh bh þ
h iT T
h¼1
X κÞ, κij ¼
T 1
2 yij1 r; . . . ; yijnij r , κi ¼ κi1T ; . . . ; κiJT , and κ ¼ κ1T ; . . . ; κIT T ,
∗
where X , Zh and Dω are defined exactly as in Model BLOR. When theprior for β
∗ ~ ~
/constant, the posterior distribution of β is also normally distributedN β 0 ; Σ 0 , but
we set the termΣ1 2 ~ ~
0 σ β to zero in both Σ 0 and β . In Appendix B are given in details the
derivations of the full conditional distributions for the BNBR model.
The full conditional distribution of ωijt is
!
f ωijt jELSE PGðyijt þ r, xTi β∗ þ b1j þ b2ij Þ ð4:16Þ
The full conditional distribution of bh, with h ¼ 1 , 2, is given as

f ðbh jELSEÞ N b~ h ; Fh ð4:17Þ
1
If η1 ¼ X β∗ + Z2b2, then F1 ¼ σ 2 1
b1 G1 þ Z1 Dω Z1
T
, ~b 1 ¼
T
F1 Z1 κ Z1T Dω η1 , and then b1 j ELSE N ~b 1 ; F1 . By defining η ¼ X β∗
2
+ Z1b1 in a similar way, we arrive at the full conditional of b2 as

1
b2 j ELSE N ~ b 2 ; F2 , where F2 ¼ σ 2 G1 þ Z T Dω Z2
b2 , bb 2 ¼ F2 Z T κ
2 2 2
Z2T Dω η2 Þ.
The full conditional distribution of σ 2bh is

f σ 2bh jELSE χ 2 ~ν b ¼ νbh þ nbh ; ~
S b ¼ bhT G1
h bh þ νbh Sbh =νbh þ nbh
ð4:18Þ
The conditional distribution of σ 2β∗ is
f ðσ 2β∗ jELSEÞ
¼ νβ∗ þ I, Sβ

χ 2 νβ∗
¼ ½ðβ∗ β0 ÞT Σ1 ∗
0 ðβ β0 Þ þ νβ∗ Sβ∗ =νβ∗ þ IÞ ð4:19Þ
Taking advantage of the fact that the NB distribution can also be generated
X Lusing
a Poisson representation as pointed out by Quenouille (1949) as Y ¼ u,
l¼1 l
where ul Logðπ Þ and is independent of L ~ Pois(r log(1 π)), where Log and
Pois denote logarithmic and Poisson distributions, respectively. Then we infer a
latent count L for each Y NB(μ, r) conditional on Y and r. Therefore, following
Zhou et al. (2012), we obtain the full conditional of r by alternating
!
I X
X J X
nij 1
f ðrjELSEÞ G a0 log 1 π ij ; PI PJ Pnij
i¼1 j¼1
t¼1 b0 þ i¼1 j¼1 t¼1 Lijt
ð4:20Þ

f Lijt jELSE CRT yijt ; r ð4:21Þ
where CRT(yijt, r) denotes a Chinese restaurant table (CRT) random

count variable
that
yijt r
can be generated as Lijt ¼ Σ l¼1 dl , where dl Bernoulli and π ij ¼
l1þr
expðη∗
ij Þ
1þexpðη∗
.
ij Þ
4.2.5 Gibbs Sampler for Model BNBR
The Gibbs sampler for the latent parameters of model BNBR with G E can be
implemented by sampling repeatedly from the following loop:
Step 1. Sample ωijt values from the Pólya-Gamma distribution in (4.16).
Step 2. Sample Lijt ~ CRT(yijt, r) from (4.21)
Step 3. Sample the scale parameter (r) from the gamma distribution in (4.20).
Step 4. Sample the location effects ( β∗) from the normal distribution in (4.15).
Step 5. Sample the random effects (bh) with h ¼ 1 , 2, from the normal distribution
in (4.17).
Step 6. Sample the variance effect (σ 2bh Þ with h ¼ 1 , 2, from the scaled inverted χ 2
distribution in (4.18).
Step 7. Sample the variance effect (σ 2β∗ Þ from the scaled inverted χ 2 distribution in
(4.19).
Step 8. Return to step 1 or terminate when chain length is adequate to meet
convergence diagnostics.
4.2.5.1 Simulation Examples
We performed a simulation study under linear predictor (Equation 4.1) with two
scenarios (S1 and S2) to show the performance of the proposed Gibbs sampler for
ordinal categorical and count phenotypes that takes into account G E. Scenario N
1 has three environments (I ¼ 3), 20 genotypes (J ¼ 20), G1 ¼ I20, G2 ¼ I3 G1,
and σ b1 ¼ σ b2 ¼ 0:5, with four different numbers of replicates of each genotype in
2 2
each environment, nij ¼ 5 , 10 , 20 , and 40. Scenario 2 is equal to scenario 1, except

that G1 ¼ 0.7I20 þ 0.3J20, where J20 is a square matrix of ones of order 20 20; this
second scenario was done with the intention of mimicking the correlation between
lines of real data available in genomic selection. We computed 20,000 MCMC
samples. Bayes estimates were computed using 10,000 samples because the first
10,000 were discarded as burn-in, and we performed 50 replications for each
scenario in both models. Next we provide the details of the simulation under each
model.
4.2.6 Model BLOR
Under the model BLOR, we simulated data from the following liability:
lijt ¼ xTi β þ b1j þ b2ij þ εijt
since i ¼ 1 , 2 , 3, j ¼ 1 , . . . , 20, and t ¼ 1 , . . . , nij, βT ¼ [6, 5, 7], and the vectors

xTi ¼ ½xi1 , xi2 , xi3 , where xi1 , xi2 , and xi3 are indicator variables that take the value
of 1 if the observed environment i is 1, 2, or 3, respectively, and 0 otherwise. The
threshold parameters used were γ 1¼-3, γ 2 ¼ 1, γ 3 ¼ 1, and γ 4 ¼ 3. The error terms
(εijt) were obtained from an L(0, 1). Then the response variable was generated as
8
>
> 1 if 1 < lij < γ 1 ,
<
2 if γ 1 < lij < γ 2 ,
yij ¼
>
> ⋮
:
5 if γ 4 < lij < 1
The priors used were not informative for

2 T
f βσ β N β0 ¼ ½0; 0; 0; I3 10000 , f ðb1 jσ b1 Þ Nð0, G1 σ b1 Þ, a normal prior
2 2
distribution for f ðb2 jσ 2b2 Þ Nð0, G2 σ 2b2 Þ, with G1 and G2 as defined above for
scenarios 1 and 2, and for the hyper-parameters for thresholds, we used γ min ¼ 4
and γ max ¼ 4. Results of this simulation study are given in Table 4.1.
4.2.7 Model BNBR
Also in this model, the priors used for both scenarios (S1 and S2) in the simulation study
were not informative for all parameters: for

f β∗ σ 2β∗ N β0T ¼ ½0; 0; 0; I3 10000 , for f(r) ~ G(0.001,1/0.001), for σ 2b1 and
σ 2b2 a ~χ 2(0.50002,4.0002), while for f ðb1 jσ 2b1 Þ N J ð0, G1 σ 2b1 Þ and
f ðb2 jσ b2 Þ N IJ ð0, G2 σ b2 Þ, with G1 and G2 as defined above for scenarios S1 and S2.
2 2
We report average estimates obtained by using the proposed Gibbs sampler along with
standard deviations (SD), which are given in Table 4.2.
4.2.7.1 Model Implementation
The Gibbs samplers described above (for model BLOR and model BNBR) were
implemented using the R-software (R Core Team 2015). For the implementation of
the proposed models, we used MCMC with the Gibbs sampler algorithm (Gelfand
and Smith 1990). We performed a total of 60,000 iterations, and 30,000 samples were
used for inference since 30,000 were used as burn-in. Thinning of the chains was not
applied following the suggestions of Geyer (1992), MacEachern and Berliner (1994),
and Link and Eaton (Link and Eaton 2012). For the implementation of model BLOR
for the real data sets, we used the following hyper-parameters νβ ¼ 3,
Sβ ¼ 0:001, β0T ¼ ½0; 0; 0, Σ0 ¼ I3 10000, γ min ¼ 1000, and γ max ¼ 1000 for
threshold parameters; these hyper-parameters lead weakly informative
! priors. For
model BNBR, the priors used were f β∗ jσ 2β∗ N p ðβ0 ¼ 03T , I3
10, 000Þ; f ðb1 jσ 2b1 Þ N J ð0nb1

T
, G1 σ 2b1 Þ, where G1 is the GRM, that is, the covariance
matrix ! of lines; f σ 2b1 χ 2 ðνb1 ¼ 3; Sb1 ¼ 0:001Þ;
f b2 jσ 2b2 N IJ ð0nb2
T
, G2 σ 2b2 Þ, G2 is the covariance matrix that belong to the
Table 4.1 Average values (mean) and standard deviation (SD) of model BLOR with four sample sizes (nij)
nij ¼ 5 nij ¼ 10 nij ¼ 20 nij ¼ 40
S Parameter True Mean SD Mean SD Mean SD Mean SD
β0 0.5 0.235 1.633 0.352 0.841 0.526 0.387 0.531 0.293
β1 1 1.033 0.324 0.952 0.298 1.00 0.287 1.00 0.239
1 β2 0.5 0.576 0.262 0.529 0.305 0.492 0.273 0.478 0.273
γ1 3 2.858 1.567 2.933 0.823 3.032 0.337 3.069 0.137
γ2 1 0.806 1.583 0.911 0.796 1.034 0.324 1.052 0.145
γ3 1 1.255 1.579 1.109 0.781 0.959 0.316 0.954 0.147
γ4 3 3.251 1.607 3.145 0.779 2.990 0.344 2.962 0.177
σ 2b1 0.5 0.587 0.203 0.601 0.202 0.548 0.137 0.568 0.155
σ 2b2 0.5 0.542 0.182 0.528 0.162 0.506 0.138 0.5437 0.133
β0 0.5 0.433 1.302 0.506 0.893 0.417 0.619 0.625 0.585
β1 1 1.222 0.350 1.089 0.272 0.956 0.222 0.968 0.224
2 β2 0.5 0.459 0.346 0.511 0.261 0.501 0.239 0.499 0.186
γ1 3 2.959 1.245 3.089 0.744 2.919 0.334 3.047 0.212
γ2 1 0.880 1.242 1.044 0.728 0.8963 0.326 1.035 0.205
γ3 1 1.2422 1.259 1.000 0.753 1.111 0.335 0.978 0.209
γ4 3 3.424 1.299 3.081 0.793 3.135 0.366 2.964 0.247
4 Bayesian Genomic-Enabled Prediction Models for Ordinal and Count Data
σ 2b1 5 0.577 0.157 0.52 0.194 0.519 0.218 0.519 0.184

σ 2b2 0.5 0.510 0.149 0.433 0.105 0.407 0.103 0.416 0.105
S denotes scenario
73
74
Table 4.2 Average values (mean) and standard deviation (SD) of model BNBR with four sample sizes (nij)
nij ¼ 5 nij ¼ 10 nij ¼ 20 nij ¼ 40
S Parameter True Mean SD Mean SD Mean SD Mean SD
β0 1.5 1.484 0.357 1.488 0.269 1.542 0.233 1.549 0.213
β1 1 0.981 0.256 0.994 0.247 1.075 0.250 1.016 0.190
1 β2 1 0.997 0.270 0.985 0.223 0.994 0.268 0.949 0.223
r 5 5.079 0.916 5.078 0.519 5.017 0.471 5.027 0.330
σ 2b1 0.5 0.542 0.196 0.594 0.176 0.582 0.180 0.590 0.216
σ 2b2 0.5 0.503 0.134 0.524 0.136 0.531 0.110 0.512 0.114
β0 1.5 1.4808 0.5009 1.4596 0.5041 1.5611 0.6108 1.4723 0.4979
β1 1 1.0631 0.2348 0.9975 0.2040 1.008 0.2226 1.025 0.1908
2 β2 1 0.9504 0.2356 1.0294 0.2167 0.9925 0.1954 0.9685 0.2018
r 5 5.1030 0.8060 4.9901 0.5928 5.0367 0.3485 5.0275 0.2033
σ 2b1 0.5 0.5422 0.1827 0.5650 0.2199 0.5785 0.1872 0.5296 0.1837
σ 2b2 0.5 0.4987 0.1155 0.5084 0.1423 0.5302 0.1301 0.5123 0.1047
S denotes scenario
O.A. Montesinos-López et al.

G E term; f σ 2b2 χ 2 ðνb2 ¼ 3; Sb2 ¼ 0:001Þ; and f(r) ~ G(a0 ¼ 0.01,1/
(b0 ¼ 0.01)). These hyper-parameters produces weakly informative priors.
4.2.7.2 Measures of Predictive Performance
Scoring Rules
Here we present the scoring rules (scoring functions) for assessing prediction
accuracy for ordinal categorical and count data. A scoring rule provides a summary
measure for evaluating probabilistic predictions by assigning a numerical score
based on the predictive distribution on the value or event that materializes
(Garthwaite et al. 2005). Assuming that with both prediction models (for ordinal
and count) we return a predictive distribution P for each observed outcome ( y) in
the data set, then we can use a scoring function to give a reward of s(P, y) if the kth
event occurs. We write s(P, Q) for the expected value of s(P, .) under Q. When a
proper scoring rule is implemented, the highest expected reward is obtained by
reporting the true probability distribution (Czado et al. 2009). Proper scoring rules
always encourage the forecaster to be honest and maximize the expected reward.
Usually the mean score is reported, which can be expressed as
1X n
S¼ s PðkÞ ; yðkÞ ,
n k¼1
where y(k) and P(k) denote the kth observed outcome and the kth predictive
distribution.
Proper Scoring Rules
A scoring rule is strictly proper if it is uniquely optimized by true probabilities.

Suppose that our best prediction with our model is the predictive distribution Q. Let
us assume that our prediction model has no incentive to predict any P 6¼ Q and is
encouraged to quote its true belief, P ¼ Q, if
sðQ; QÞ sðP; QÞ,
with equality if, and only if, P ¼ Q. A scoring rule with this characteristic is said to
be strictly proper if s(Q, Q) s(P, Q) for all P and Q. Propriety is an essential
property of a scoring rule that encourages coherent and honest predictions. Strict
propriety ensures that both calibration and sharpness are being addressed (Czado
et al. 2009), understanding sharpness as the concentration of the predictive distri-
bution, and the shorter the sharper the predictions and the sharper the better, subject
to calibration.
Examples of Proper Scoring Rules
The logarithmic score is defined as
logsðP; yÞ ¼ logpy
where py refers to the probability mass function at the observed outcomes (Czado
et al. 2009). This is the only proper scoring rule that depends on the predictive
distribution P through its probability mass ( py). The quadratic score or Brier score
is defined as
qsðP; yÞ ¼ 2py þ k p k2 , ð4:22Þ
X
1
where k p k2 ¼ p2k , which can frequently be computed analytically for the
k¼1
Poisson and the negative binomial distribution. The spherical score is defined as
py
sphsðP; yÞ ¼
kpk
The ranked probability score (Czado et al. 2009) was originally proposed for
ranked categorical data. It is defined as
X
1
rpsðP; yÞ ¼ fPl 1ðy lÞg2 ,
l¼1
In practice it is not easy to choose a scoring rule, unless there is a unique and
clearly defined underlying decision problem. However, it is always preferable to
use a proper scoring rule instead of a classical measures of predictive performance
that is not proper, such as the mean absolute error, mean square error of prediction,
and mean squared Pearson residuals. However, in many situations, probabilistic
predictions have multiple simultaneous uses, and it may be appropriate to use a
variety of tools and scores, to take advantage of their different emphases and
strengths (Czado et al. 2009). For example, Wecker (1989) used the Brier score
(quadratic score) in the assessment of time series predictions of counts.
Montesinos-López et al. (2015a, 2015b) also used the Brier score for assessing
prediction accuracy for ordinal data in the context of genomic-enabled prediction.
4.2.7.3 Assessing Prediction Accuracy
Following Burgue~ no et al. (2012), we implemented the cross-validation scheme

(CV2). Cross-validations mimic real situations that a breeder might face. This
cross-validation scheme (CV2) mimics a situation where lines were evaluated in
some environments but missing in others. In this case, information from relatives is
used, and prediction assessment can benefit from borrowing information between
lines within an environment and between lines across environments.
For this cross-validation, we used tenfold cross-validations; each time, ninefolds
were used for training and onefold for testing. The training set was used to fit the
model, and the validation set was used to evaluate the prediction accuracy of the
proposed models. However, only the information of 10% of the lines in one
environment was missing. Among the variety of methods for assessing prediction
accuracy for ordinal categorical and count data, we used the Brier score for the GLS
data set and the Spearman correlation (Cor) and mean square error of prediction
(MSEP) for the FHB data set. These two criteria were used to obtain comparable
predictions between the proposed models for counts (BNBR and Poisson) and
models for normal and lognormal data. However, we need to point out that proper
scoring rules (as the Brier score) should be preferred since classical measure of
predictive performance as the MSEP and Cor are not proper and are not the best
option for ordinal or counts data. For the GLS data sets with ordinal categorical
data, the Brier score (Brier 1950) was computed as
X
n X
C 2
BS ¼ n1 πbkc dkc ð4:23Þ
k¼1 c¼1
where BS denotes the Brier score, πbk denotes the predictive distribution derived
from the estimated model for observation k, and dkc takes a value of 1 if the ordinal
categorical response observed for individual k falls into category c; otherwise,
dkc ¼ 0. The range of BS in Eq. (4.23) for ordinal data is between 0 and 2. For
this reason, we divided BS/2 to get the Brier score bounded between 0 and 1; lower
scores imply better predictions. For count data, the Brier score was computed using
Eq. (4.22) but computing py and kpk2 depending on the model used, e.g., with the
Poisson probability mass function if the Pois model was used for fitting the data, or
the negative binomial probability mass function if the BNBR model was used.
Here, the Brier score can be any real value, and the lower the value, the better the
model.
The Brier score rule uses all the information contained in the predictive distri-
bution, not just a small portion like the hit rate or the log likelihood score.
Therefore, it is a reasonable choice for comparing ordinal categorical and count
regression models, although there are other scoring rules that also have good
properties.
4.3 Results
In the following sections, we investigate the performance of the proposed BLOR

and BNBR models through a simulation study and with real data.
4.3.1 Simulated Data Sets
The only purpose of the simulations performed in this section is to show that the
proposed methods work well in terms of parameter estimation.
4.3.1.1 Model BLOR
In Table 4.1, we report average estimates obtained by all methods, along with
standard deviations (SD) for both scenarios (S1 and S2) under study. Table 4.1
shows that the bias in the estimation of the parameters is a little larger in S1
compared to S2 (which takes into account the GRM). Also, parameter β1 is the
parameter with larger bias (underestimated) when the sample size is nij ¼ 5 and 10.
Both variances ( σ 2b1 , σ 2b2 ) are overestimated under scenario 1, but only σ 2b1 is
overestimated under scenario 2. From Table 4.1, it is clear that as the sample size
increases, the average biases and SD decrease in all cases. This confirmed the
consistent properties of all the estimates.
4.3.1.2 Model BNBR
Table 4.2 gives the results of the simulation study under both scenarios (S1 and S2).
Again, the bias in the estimation of the parameters is a little larger in S1 compared
to S2. Table 4.2 shows that parameter β0 is the parameter with larger bias
(underestimated). Both variances ( σ 2b1 , σ 2b2 ) are overestimated under scenario
1, but only σ 2b1 is overestimated under scenario 2. Also, with sample size nij ¼ 5,
the parameter r shows the larger SD; however, for larger sample sizes (nij ¼ 20 , 40),
the SD are considerably reduced. In general, there is no large reduction in SD when
the sample size increases from 5 to 10, 20, and 40, the exception being the
estimation of r under both scenarios and the estimation of β0 under scenario
1, where there is a large reduction of SD when the sample size increases. Even
though the estimations under both models are not perfect, the proposed Gibbs
samplers for ordinal and count data that take into account G E do a good job
for estimating the parameters since the estimates are close to the true values and
with a SD of reasonable size. However, in both models, a more in-depth simulation
study is required to ensure that these findings are valid for all possible scenarios.
4.3.2 Real Data Sets
Using the real data sets, we compared four scenarios for each model (Table 4.3).
The table shows that scenarios S1 and S2 do not take into account interaction effects
in the linear predictor, only main effects. Also, scenarios S1 and S3 do not use
Table 4.3 Scenarios Main effects Interaction effects

proposed to fit the real data set
Scenario E L G EL EG
with both models
S1 X X
S2 X X
S3 X X X
S4 X X X
E denotes environments, L lines, G lines with markers, EL inter-
action effect between environment and line, and EG interaction
effect between environments and lines with markers
Table 4.4 Posterior average values (mean) and standard deviation (SD) for the GLS data set with
model BLOR for each scenario given in Table 4.3
S1 S2 S3 S4
Parameter Mean SD Mean SD Mean SD Mean SD
β1 0.422 0.063 0.217 0.069 0.324 0.108 0.415 0.093
β2 0.167 0.056 0.376 0.064 0.524 0.099 0.342 0.089
β3 0.254 0.070 0.047 0.075 0.042 0.111 0.273 0.094
γ1 1.625 0.053 1.430 0.061 2.084 0.076 2.200 0.070
γ2 0.434 0.041 0.227 0.056 0.278 0.057 0.447 0.039
γ3 0.546 0.041 0.751 0.064 1.105 0.061 0.905 0.050
γ4 1.355 0.047 1.566 0.070 2.182 0.072 1.964 0.070
σ 2b1 0.205 0.030 0.200 0.032 0.097 0.075 0.321 0.072
σ 2b2 1.488 0.144 1.525 0.164
postMeanLogLik 3489 3482 2729 2769
marker information. These four scenarios are studied with the goal of investigating
the gain in model fit and prediction ability taking into account the interaction effect
and using the marker information available.
4.3.3 Model BLOR for GLS Data Set
Table 4.4 depicts the posterior mean and posterior standard deviation of the
parameter estimates with the model BLOR using the GLS data set. The parameter
estimates and posterior mean log likelihood (postMeanLogLik) of S1 and S2 are
alike, as are the parameter estimates and posterior mean log likelihood
(postMeanLogLik) of S3 and S4. The postMeanLogLik favors models S3 and S4.
In Table 4.5, we see that in Colombia, the best model in terms of prediction
accuracy using the Brier score was S4, while the worst was S1. In Harare, the best
model was S3 and the worst was S1. In Mexico, the best model was S2 followed by
S4, while the worst model was S3. It is important to point out that prediction
accuracy were best in Mexico and worst in Colombia. Because the differences
between scenarios in each country are not large, it is not easy to discriminate
Table 4.5 Brier scores (mean and standard deviation (SD)) from model BLOR evaluated for
validation samples for each scenario given in Table 4.3 for the GLS data set. Lower scores indicate
better predictions
Colombia Harare Mexico
Scenario Mean SD Mean SD Mean SD
S1 0.417 0.020 0.388 0.016 0.360 0.044
S2 0.407 0.015 0.382 0.019 0.353 0.039
S3 0.404 0.023 0.363 0.015 0.367 0.035
S4 0.399 0.021 0.383 0.025 0.359 0.038
Table 4.6 Estimated fixed effects, variance components, and deviance information criteria (DIC)
for models BNBR and Pois for the GLS data set
S1 S2 S3 S4
Parameter Mean SD Mean SD Mean SD Mean SD
Model BNBR
β∗1 4.002 0.071 4.007 0.062 4.001 0.055 4.003 0.045
β∗2 3.828 0.074 3.833 0.064 3.822 0.057 3.812 0.05
β∗3 3.961 0.075 3.966 0.066 3.957 0.059 3.954 0.053
r 143.656 9.771 144.249 8.745 141.124 7.339 139.77 6.627
σ 2b1 0.032 0.004 0.034 0.005 0.033 0.005 0.037 0.005
σ 2b2 – – – – 0.034 0.004 0.0368 0.005
DIC 8516.966 (3) 8484.151 (2) 8564.722 (4) 8462.392 (1)
Model Pois
β∗1 5.952 0.026 5.95 0.022 5.968 0.029 5.982 0.024
β∗2 5.772 0.019 5.772 0.015 5.785 0.022 5.781 0.016
β∗3 5.907 0.03 5.909 0.029 5.924 0.034 5.925 0.029
r 1000 _ 1000 – 1000 – 1000 –
σ 2b1 0.033 0.004 0.035 0.005 0.034 0.005 0.037 0.005
σ 2b2 – – – – 0.034 0.004 0.037 0.004
DIC 8488.566 (3) 8457.023 (2) 8533.377 (4) 8427.161 (1)
() denotes the ranking of the four scenarios with the DIC for each model
between models; even though in two of the three locations, scenario S4 was
identified as the best model for prediction.
4.3.4 Model BNBR for GLS Data Set
The posterior means, posterior SD of the scalar parameters, and deviance informa-
tion criteria (DIC) for model BNBR and model Pois are given in Table 4.6. The
table shows that the posterior means of the beta regression coefficients (β∗ ∗
1 , β2 , and
a b c
8
10
8
8
6
6
Density
Density
Density
6
4
4
4
2
2
2
0
0
–4.1 –3.9 –3.9 –3.7 –4.1 –3.9 –3.7
– – –
β1 β2 β3
d e f
70
70
0.06
50
50
0.04
Density
Density
Density
30
30
0.02
0 10
0 10
0.00
120 140 0.02 0.04 0.06 0.02 0.04 0.06

2 2
r σb1 σb2
Fig. 4.2 Histogram representation of posterior distributions of scalar parameters for scenario S4
and model BNBR fitted with the whole data set GLS for ðaÞ β∗ ∗ ∗
1 , ðbÞ β 2 , ðcÞ β 3 , ð dÞr, ðeÞ σ b1 , and
2
ðfÞ σ b2 with priors superimposed as dashed lines at the bottom

2
β∗
3 Þ, variance components and scale parameter (r) were similar for the four pro-
posed scenarios for model BNBR with the exception of scenarios S3 and S4, where
the parameter r was lower, while the posterior SD of scenarios S3 and S4 were
lower than those of scenarios S1 and S2. With regard to the DIC in model BNBR,
the scenarios rank as follows: rank 1 for scenario S4, rank 2 for scenario S2, rank
3 for scenario S1, and rank 4 for scenario S3. Figure 4.2 shows a histogram
representation of the posterior distributions for scalar parameters, and in all plots,
the priors for each parameter in model BNBR are not informative.
Table 4.6 also shows that the posterior means and posterior SD of the beta
regression coefficientsðβ∗ ∗ ∗
1 , β2 , and β3 Þ and variance components for model Pois
are similar between the four proposed scenarios. In terms of DIC, the scenarios rank
as follows: rank 1 for scenario S4, rank 2 for scenario S2, rank 3 for scenario S1,
a b c
14
20 25
15
6 8 10
Density
Density
Density
10
10 15
5
4
5
0 2
0
–6.05 –5.95 0 –5.84 –5.78 –5.72 –6.00 –5.90 –5.80

– – –
β1 β2 β3
d e
70
70
50
50
Density
Density
30
30
0 10
0 10
0.02 0.04 0.06 0.02 0.04 0.06

2
2
σb1 σb2
Fig. 4.3 Histogram representation of posterior distributions of scalar parameters for scenario
4 and model Pois fitted with the whole data set GLS for ðaÞ β∗ ∗ ∗
1 , ðbÞ β 2 , ðcÞ β 3 , ðdÞ σ b1 , and ðeÞ σ b2
2 2
with priors superimposed as dashed lines at the bottom
and rank 4 for scenario S3. Figure 4.3 also shows that the priors for each scalar
parameter in model Pois are not informative.
Table 4.6 indicates that model Pois fits this real data set best, since comparing
the DIC for these two models there is a clear superiority of model Pois over model
BNBR. This means that model Pois is enough for this data set.
In Table 4.7 we present the mean and SD of the Brier scores resulting from the
tenfold cross-validation performed. The Brier scores given in Table 4.7 were
calculated using the testing set. According to the Brier scores, in model BNBR,
the best model for prediction was S3 in the three environments. Under model Pois,
the scenario with the best prediction accuracy was scenario S3 for Harare and
Mexico, but scenario S2 in Colombia. It is not clear which model is the best (model
Pois or model BNBR) since the Brier scores are very similar for both models. This
may be due to the fact that the data are not really count, because we used the GLS
data set that has the response variable as categorical ordinal.
Table 4.7 Brier scores (mean and standard deviation (SD)) from the BNBR and Pois models
evaluated for validation samples for each scenario given in Table 4.3 for the GLS data set
Colombia Harare Mexico
Model Mean SD Mean SD Mean SD
BNBR model
S1 0.224 0.013 0.232 0.013 0.219 0.023
S2 0.226 0.010 0.229 0.012 0.219 0.020
S3 0.233 0.011 0.242 0.011 0.233 0.014
S4 0.221 0.008 0.225 0.011 0.219 0.017
Pois model
S1 0.228 0.019 0.216 0.027 0.264 0.028
S2 0.231 0.018 0.218 0.020 0.248 0.028
S3 0.194 0.174 0.228 0.036 0.281 0.025
S4 0.208 0.015 0.218 0.016 0.234 0.025
Lower scores indicate better predictions
4.3.5 Model BNBR for FHB Data Set
These results were taken from the paper of Montesinos-López et al. (2016). For this
data set (FHB data) was implemented with the following linear predictor:
ηijk ¼ Ei + R(E)ik + gj + gEij with i ¼ 1 , . . . , I ; j ¼ 1 , 2 , . . . , J , k ¼ 1 , . . . , K ,
t ¼ 1 , 2 , . . . , nijk, where Ei represents environment i, R(E)ik represents the effect
of block k within environment i, gj is the marker effect of genotype j, and gEij is the
interaction between markers and the environment. This predictor is very similar to
that given in Eq. (4.1) with the term R(E)ik added. We assume that yijkt represents
the count response for the tth replication of the jth line in the kth block in the ith
environment. Four models were implemented using the linear predictor given
above: model BNBR, model Poisson, normal, and lognormal (LN). Also for each
model, the four scenarios given in Table 4.3 were studied. These models were
implemented under a Bayesian approach (these Bayesian models were
implemented through Gibbs sampler since all full conditionals were derived ana-
lytically, for the BNBR and Poisson model the Pólya-Gamma augmentation
approach explained above was used, see Montesinos-López et al. Montesinos-
López et al. 2015a, 2015b, 2015c and Montesinos-López et al. 2016 for details of
all derivations). The Gibbs samplers for the four models were implemented in R
Core Team ( 2015) using 60,000 iterations with a burn-in of 30,000, so that 30,000
samples were used for inference. The prediction accuracy of the proposed models
were evaluated with the Spearman correlation (Cor) and the mean square error of
prediction (MSEP), both calculated using the observed and predicted response
variables of the validation set resulting of a tenfold cross-validation implemented
for the four models and four scenarios. In this example we used the Spearman
correlation and the MSEP to compare the results with models normal and
lognormal.
In Table 4.8 we can see that the ranking of scenarios for the BNBR model were
as follows: in Batan 2012, 1 for S4, 2 for S3, 3 for S1, and 4 for S2. In Batan 2014,
the ranking was 1 for S4, 2 for S3, and 3 for S1 and S2. In Ecuador 2014, the ranking
was 1 for S3, 2 for S2, 3 for S1, and 4 for S4. With the MSEP, the ranking for model
BNBR in Batan 2012 was 1 for S3, 2 for S4, and 3 for S1 and S2. In Batan 2014, the
ranking was 1 for S2, 2 for S1, 3 for S3, and 4 for S4. In Ecuador 2014, the ranking
in terms of MSEP was 1 for S3, 2 for S2, 3 for S4, and 4 for S1. Under model Pois,
the ranking of the four scenarios in each locality was exactly the same as the
ranking reported for model BNBR. For model normal in terms of the Spearman
correlation, S1 was the best in prediction accuracy in Batan 2012, while S4 was the
worst in all three locations. In terms of MSEP, the best scenario was S3 in Batan
2012 and Ecuador 2014, and the worst was S4 in Batan 2014 and Ecuador 2014. For
model LN in terms of the Spearman correlation, the best scenarios were scenarios
S1, S2, and S3, and the worst was S4 in Batan 2012. In Batan 2014, the best
scenario was S1, then scenario S3, and the worst was scenario S4. In Ecuador 2014,
the best scenario was scenario S1 and S3, then S2 and S4. In terms of MSEP for
Batan 2012, the best scenario was S3, then S1 and S2, and the worst was S4. In
Batan 2014, the best scenario was S1, then S2, and the worst was scenario S4.
Finally, in Ecuador 2014, the best scenario was S3, then S2, and the worst was
scenario S1.
Table 4.9 gives the average of the ranks of the two posterior predictive checks
(Cor and MSEP) that were used. Since we are comparing four scenarios for each
model, the values of the ranks range from 1 to 4, and the lower the values, the better
the scenario. For ties we assigned the average of the ranges that would have been
assigned had there been no ties. Table 4.9 shows that the best scenarios were
scenarios S3 and S4 under models BNBR and Pois in Batan 2012. In Batan
2014, under models BNBR and Pois, the best scenario was S2, while in Ecuador
2014, the best scenario was S3. Under model normal, the best scenario was S1 in
Batan 2014, S1 and S3 in Ecuador 2014, while in Batan 2012, the best scenarios
were S2 and S3. Finally, under model LN, the best scenario was S3 in Ecuador
2014, S3 in Batan 2012, and S1 in Batan 2014. Then according with results of
Tables 4.8 and 4.9, the best models in terms of prediction accuracy are models
BNBR and Pois, since they had better predictions in the validation set based on
both the posterior predictive checks (Cor and MSEP) implemented. Also, we
observed that in models BNBR and Pois, taking into account G E considerably
increased the prediction accuracy, which was expected since there is enough
scientific evidence that including G E interaction improves prediction accuracy.
However, to use these models correctly, it is important to first understand the types
of data we have before deciding on the modeling approach to be used.
Table 4.8 Estimated posterior predictive checks with cross-validation for models BNBR, Pois,
normal, and LN for the FHB data set
Batan 2012 Batan 2014 Ecuador 2014
Scenario Cor MSEP Cor MSEP Cor MSEP
Model BNBR
S1 Mean 0.43 (3) 0.98 (3.5) 0.43 (3.5) 1.39 (2) 0.18 (3) 11.733 (4)
SD 0.33 0.72 0.33 1.35 0.40 9.471
S2 Mean 0.42 (4) 0.98 (3.5) 0.43 (3.5) 1.38 (1) 0.20 (2) 11.222 (2)
SD 0.33 0.72 0.33 1.36 0.37 8.614
S3 Mean 0.54 (2) 0.49 (1) 0.52 (2) 1.48 (3) 0.22 (1) 8.645 (1)
SD 0.28 0.38 0.29 2.32 0.39 5.688
S4 Mean 0.56 (1) 0.61 (2) 0.56 (1) 1.85 (4) 0.12 (4) 11.343 (3)
SD 0.24 0.44 0.22 2.68 0.41 8.154
Model Pois
S1 Mean 0.43 (3) 0.98 (3.5) 0.43 (3.5) 1.39 (2) 0.18 (3) 11.733 (4)
SD 0.33 0.72 0.33 1.35 0.40 9.471
S2 Mean 0.42 (4) 0.98 (3.5) 0.43 (3.5) 1.38 (1) 0.20 (2) 11.222 (2)
SD 0.33 0.72 0.33 1.36 0.37 8.614
S3 Mean 0.54 (2) 0.48 (1) 0.52 (2) 1.48 (3) 0.22 (1) 8.645 (1)
SD 0.28 0.38 0.29 2.32 0.39 5.688
S4 Mean 0.56 (1) 0.61 (2) 0.56 (1) 1.85 (4) 0.12 (4) 11.343 (3)
SD 0.24 0.44 0.22 2.68 0.41 8.154
Model normal
S1 Mean 0.36(1) 1.10 (4) 0.37 (1.5) 1.79 (1) 0.15 (1.5) 7.425 (2)
SD 0.28 0.88 0.39 1.70 0.32 4.151
S2 Mean 0.34 (2) 0.99 (2) 0.33 (3) 2.01 (3) 0.07 (3) 7.454 (3)
SD 0.33 0.65 0.44 2.46 0.33 4.339
S3 Mean 0.33 (3) 0.81 (1) 0.37 (1.5) 1.96 (2) 0.15 (1.5) 7.318 (1)
SD 0.30 0.46 0.40 2.99 0.29 4.159
S4 Mean 0.27 (4) 1.03 (3) 0.24 (4) 2.37 (4) 0.04 (4) 8.482 (4)
SD 0.34 0.73 0.45 3.42 0.24 4.326
Model LN
S1 Mean 0.51 (2) 0.66 (2.5) 0.46 (1) 1.60 (1) 0.15 (1.5) 8.10 (4)
SD 0.21 0.42 0.31 2.35 0.38 5.11
S2 Mean 0.51 (2) 0.66 (2.5) 0.43 (3.5) 1.78 (2) 0.09 (3.5) 7.82 (2)
SD 0.22 0.39 0.35 2.82 0.46 5.31
S3 Mean 0.51 (2) 0.64 (1) 0.45 (2) 1.871 (3) 0.15 (1.5) 7.76 (1)
SD 0.21 0.45 0.31 3.16 0.37 5.21
S4 Mean 0.43 (4) 0.72 (4) 0.43 (3.5) 1.95 (4) 0.09 (3.5) 8.04(3)
SD 0.25 0.42 0.33 3.15 0.41 5.18
The numbers in () denote the ranking of the four scenarios set for each posterior predictive check
(extracted from Montesinos-López et al. 2016)
Table 4.9 Rank averages for the four scenarios resulting from the tenfold cross-validation
implemented for the FHB data set
Batan Batan Ecuador Batan Batan Ecuador
Scenarios 2012 2014 2014 2012 2014 2014
Model BNBR Model normal
S1 3.25 2.75 3.5 2.5 1.25 1.75
S2 3.75 2.25 2 2 3 3
S3 1.5 2.5 1 2 1.75 1.75
S4 1.5 2.5 3.5 3.5 4 4
Model Pois Model LN
S1 3.25 2.75 3.5 2.25 1 2.75
S2 3.75 2.25 2 2.25 2.75 2.75
S3 1.5 2.5 1 1.5 2.5 1.25
S4 1.5 2.5 3.5 4 3.75 3.25
Each average was obtained as the mean of the rankings given in Table 4.8 for the two posterior
predictive checks (Cor and MSEP) in each scenario (extracted from Montesinos-López et al. 2016)
4.4 Conclusions
Generalized linear mixed models (GLMMs) are considered to be one of the major
methodological developments of the second half of the last century. The main
factor contributing to their wide applicability over the last 30 years or so is their
flexibility, because they can be applied to different types of data (Berridge and
Crouchley 2011), including continuous interval/scale, categorical (including binary
and ordinal) data, count data, beta data, and others. Each member of the GLMMs
family is appropriate for a specific type of data (Berridge and Crouchley 2011).
However, GLMMs for non-normal data are scarce in the context of genomic-
enabled prediction, since most of the models developed so far are linear mixed
models (mixed models for Gaussian data). For these reasons, we believe that
developing specific methods for categorical ordinal and count data for genomic-
enabled prediction can help to improve the selection of candidate genotypes early
when the phenotypes are ordinal and counts. Because using transformation to
approximate the ordinal data and counts to normality or assuming that these types
of data are normally distributed frequently produces poor parameter estimates and
lower power, parameter interpretation is more difficult when transformation is used
(Stroup 2015). However, in genomic selection, phenotypic data (dependent vari-
able) are currently not taken into account before deciding on the modeling approach
to be used, mainly due to the lack of genomic-enabled prediction models for
non-normal phenotypes. Although our proposed Bayesian regression models are
only for categorical ordinal and count data, they help fill this lack of genomic-
enabled prediction models for non-normal data.
This chapter presents recent advances in models for whole genome prediction
for ordinal categorical and count data. These models were built using the Pólya-
Gamma data augmentation approach of Scott and Pillow (2013), which produces a
Gibbs sampler with full conditional distributions similar to that of the Bayesian
probit ordinal regression (BPOR) model of Albert and Chib (1993). The proposed
Bayesian logistic ordinal regression (model BLOR) is reduced to the BPOR model
of Albert and Chib (1993) when the sampled values, ωijt, from the Pólya-Gamma
distribution in Eq. (4.6) are set to 1. This is an advantage because researchers can
perform an exact logistic or probit ordinal regression with the proposed model
without having to do approximations to perform a logistic ordinal regression. The
performance of the proposed model BLOR without interaction was compared to the
BPOR model using the approximation (logit(u) ¼ (1.75)Φ1(u)) in a small simula-
tion study and with real data sets using a 4- and 5-point ordinal scale by
Montesinos-Lopez et al. (Montesinos-López et al. 2015b). They found that the
estimation of parameters using the approximation logit(u) ¼ (1.75)Φ1(u) produces
a considerable amount of bias and can give rise to wrong conclusions in association
studies. However, they observed no differences between the two models in terms of
prediction ability with two real data sets. For this reason, the proposed BLOR
model is a viable alternative for analyzing ordinal data since it is more robust for
dealing with outlying data. This is because the logistic distribution has heavier tails
and provides regression coefficients that are more interpretable due to their con-
nection to odds ratios (Zucknick and Richardson 2014). This last advantage does
not make sense when p n,since the main driving force in Bayesian models in the
case of p n is the prior and not the data (Gianola 2013). Even with this restriction,
the proposed BLOR model unifies logistic and probit ordinal regression under a
Bayesian framework and is a useful alternative for genomic-enabled prediction of
ordinal categorical trials where available data sets have a larger number of param-
eters to estimate than observations.
The proposed Bayesian method for count data describes the work done by
Montesinos-López et al. (2016), which extended the work of Montesinos-López
et al. (2015c) to incorporate the G E term. Modeling G E for categorical ordinal
and count data in the context of genomic-enabled prediction plays a central role in
plant breeding for the selection of candidate genotypes that present high adaptation
to a wide range of environmental conditions including local conditions. Also,
incorporating G E helps to predict yet-to-be observed phenotypes when the
relative performance of genotypes varies across environments. To the best of our
knowledge, this is the first work on genomic-enabled prediction that uses the NB
and Poisson distributions with G E.
It should also be noted that to use these models correctly, it is important to first
understand the types of data being analyzed before deciding on the modeling
approach to be employed. If the phenotypic data are normally distributed, the linear
mixed models for genomic-enabled prediction developed so far for Gaussian
phenotypes should be used. If the phenotypic data are binary or categorical ordinal,
the methods proposed by Montesinos-López et al. (2015a, 2015b) and their exten-
sions given in this chapter with the logit link should be preferred. If the phenotypic
data are counts (number of panicles per plant, number of seeds per panicle, weed
count per plot, etc.) and the counts are small, the models developed in this study
(BNBR and Pois models) and those proposed by Montesinos-López et al. (2015c,
2016) are the best option, since they have more advantages over the conventional
linear mixed models with Gaussian response, as was observed when we applied
them to the real data set. However, Montesinos-López et al. (2015c, 2016) also
found that when the count response is log transformed, the prediction accuracies are
better than when using the counts as if they were normally distributed.
Finally, it is important to extend the proposed methods of Montesinos-López
et al. (2015a, 2015b, 2015c, 2016) developed under the work of Scott and Pillow
(2013) for ordered categorical responses and count data for multiple traits. Our
methods are elegant, easy to implement, and produce a unified Gibbs sampler
framework useful for both types of phenotypic responses. This is important
because, of all the computational intensive methods for fitting complex multilevel
models, the Gibbs sampler is the most popular due to its simplicity and ability to
effectively generate samples from high-dimensional probability distributions (Park
and van Dyk 2009). For this reason, we believe these methods are appealing
alternatives for plant and animal researchers. Both models can be easily extended
to take into account epistatic effects for the joint modeling of multiple traits, which,
as is well documented, can increase the prediction accuracy of the models.
Acknowledgments We would like to thank all researchers in CIMMYT’s Global Maize Program
(GMP) and Global Wheat Program (GWP), as well as the national program researchers who
generated the data used in this and other studies.
Appendix A: Derivation of Full Conditional Distributions

for Model BLOR
Liabilities and ωijt. The fully conditional posterior distribution of liability lijt is
PðljELSEÞ / Pðlj β, bÞPðyjl, γ Þ

I Y
Y J Y
nij X
C
/ f ðlijt Þ Iðyijt ¼ cÞIðγ c1 < lijt < γ c Þ
i¼1 j¼1 t¼1 c¼1
I Y
Y J Y
nij
expðlijt þ xTi β þ b1j þ b2ij Þ X
C
/ Iðyijt ¼ cÞIðγ c1 < lijt < γ c Þ
2
i¼1 j¼1 t¼1½1 þ expðlijt þ xTi β þ b1j þ b2ij Þ c¼1
ð1 " #
YI Y J Y nij
ω ðl þ x T
β þ b þ b Þ 2
22
ijt iljt 1j 2ij
/ exp i
Pðωijt ; b ¼ 2, d ¼ 0Þ
i¼1 j¼1 t¼1 0 2
X
C
dωijt Iðyijt ¼ cÞIðγ c1 < lijt < γ c Þ
c¼1
The last inequality was obtained using a technique called the Pólya-Gamma
method (Scott and Pillow 2013), which is useful when working with logistic
likelihoods, and has the form
ð1
ðeψ Þa ωψ 2
¼ 2b eκψ e 2 Pðω; b; 0Þdω
ð1 þ eψ Þb 0
where κ ¼ a b/2 and P(ω; b, d ¼ 0) denotes the density of the random variable
ω ~ PG(b, d ¼ 0), where PG(b, d) denotes a Pólya-Gamma distribution lijt with
parameters b andd and
density

b d
Pðω; b; dÞ ¼ cosh
2
b1 X
1 ð2n þ bÞ2 d 2
2 Γðn þ bÞð2n þ bÞ
ð1Þn pffiffiffiffiffiffiffiffiffiffiffi exp ω ,
ΓðbÞ n¼0 Γðn þ 1Þ 2πω3 8ω 2
where cosh denotes the hyperbolic cosine.

Then the joint posterior distribution of lijt and ωijt is equal to
" #
YI Y J Y nij
ωij ðlijt þ xTi β þ b1j þ b2ij Þ2
2
Pðl, ωjELSEÞ / 2 exp Pðωijt ; 2, 0Þ
i¼1 j¼1 t¼1
2
X
C
Iðyijt ¼ cÞIðγ c1 < lijt < γ c Þ
c¼1
Therefore, the fully conditional posterior distribution of liability lijt is a truncated

normal distribution and its density is
f ðlijt jELSEÞ
ϕ ωijt ðlijt xTi β b1j b2ij Þ
¼ pffiffiffiffiffiffiffi
Φð ωijt ðγ c xTi β b1j b2ij ÞÞ Φ ωijt ðγ c1 xTi β b1j b2ij Þ
For simplicity, ELSE is the data and the parameters, except for the one in
question. ϕ and Φ are the density and distribution function of a standard normal
random variable and the fully conditional posterior distribution lijt of ωijt is
" #
2
ω ðl þ xT
β þ b þ b Þ
f ðωijt jELSEÞ / 22 exp
ijt ijt 1j 2ij
i
Pðωijt ; 2, 0Þ
2
" #
ωijt ðlijt þ xTi β þ b1j þ b2ij Þ2
/ exp Pðωijt ; 2, 0Þ
2
From here and from Eq. (4.5) of Polson et al. (2013), we get that

f ωijt jELSE PG 2; lijt þ xiT β þ b1j þ b2ij
Regression Coefficients (β) First note that the fully conditional posterior of l , β , ω
is

Pðl, β, ωjELSEÞ / Pðlj β, b1 , b2 ÞPðyjl, γ ÞPðωÞP βσ 2β
!T !!
1 X2 X2
/ exp l þ Xβ þ Zh bh Dω l þ Xβ þ Zh bh PðωÞP βσ 2β
2 h¼1 h¼1
YI YJ Y
nij

where PðωÞ ¼ i¼1 j¼1
P ωijt ; 2; 0 . Then, the full conditional posterior
t¼1
distribution of β is
PðβjELSEÞ
!T !
1 X2 X
2
/exp lþXβþ Zh bh Dω lþXβþ Zh bh
2 h¼1 h¼1
1
ðββ0 ÞT Σ1 2
0 σβ ðββ0 ÞÞ
2
1 X2 T
/exp ½βT ðΣ1
0 σ 2
β þX T
D ω XÞβ2 Σ 1 2
0 σ β β 0 X T
D ω ð Z h b h ÞþXT
D ω l βÞ
2 h¼1
h
1 T 1 i
/ exp β ~ β0 Σ ~
0 β ~β
0
2
1 P
2
where Σ 0 ¼ ðΣ1 2
0 σ β þ X Dω XÞ ,
T
β 0 ¼ Σ 0 Σ1 2
0 σ β β0 X Dω ð
T
Zh bh Þþ
h¼1
XT Dω l . It is important to point out that if we use a prior for β / Constant
(improper uniform distribution), then in Σ ~ 0 and ~β 0 we need to make 0 the term
1 2
Σ0 σ β . Finally, the full conditional) posterior of β is

βjELSE N I β~ ;Σ
~0
0
Polygenic effects (bh) Now the full conditional posterior of bh is given as

Lðbh jELSEÞ !
1 X 2
T
X2
/ exp ðl þ Xβ þ Zh bh Þ Dω ðl þ Xβ þ Zh bh Þ Pðbh jσ 2bh Þ
2 h¼1 h¼1

1 h T 2 1 T i
/ exp bh σ b G þ Zh Dω Zh bh 2 Zh Dω l Zh Dω Xβ bh
T T T
2

1 T
/ exp bh ~ b h F1
h b h ~
b h
2
This implies that the full conditional posterior of bh is

1 T 1
f ðbh jELSEÞ N bh ¼ Fh ðZhT Dω l ZhT Dω ηh Þ, Fh ¼ ðσ 2
bh G h þ Z h D ω Z h Þ
T
with h ¼ 1 , 2, η1 ¼ Xβ + Z2b2 and η2 ¼ Xβ + Z1b1.

Variance of polygenic effects ðσ 2bh Þ: Next, the conditional distribution of σ 2bh is
obtained. If σ 2bh χ 2 ðνh , Sh Þðshape and scaleÞ, then
!
T 1
1 b G b hh þ ν h S h
Pðσ 2bh jELSEÞ / νh þnh exp h h
ðσ 2 Þ 2 þ1 2σ 2bh
bh
This is the kernel of the scaled inverted χ 2 distribution; therefore, the full
conditional posterior is

f ðσ 2bh jELSEÞ χ 2 νh ¼ νh þ nh , Sb ¼ ðbhT G1
h bh þ νh Sh Þ=νb þ nh Þ
Threshold effects (γ) The density of the full conditional posterior distribution of
the cth threshold, γ c, is
PðγjELSEÞ / Pðyjl, γÞPðγ Þ
I Y
Y J Y
nij X
C
/ Iðyijt ¼ cÞIðγ c1 < lijt < γ c ÞIðγ 2 TÞ ð4:A:1Þ
i¼1 j¼1 t¼1 c¼1
If Eq. (4.A.1) is seen as a function of γ c, it is evident that the value of γ c must be

larger than all the lijt|yijt ¼ c and smaller than all the lijt|yijt ¼ c þ 1. Hence, as a
function of γ c, Eq. (4.A.1) leads to the uniform density
1
Pðγ c jELSEÞ ¼ I ðγ 2 TÞ ð4:A:2Þ
min lijt yijt ¼ c þ 1 max lijt yijt ¼ c

Equation (4.A.2) corresponds to a uniform distribution on the interval [min

{min (lijt|yijt ¼ c þ 1), γ c þ 1, γ max }, max{max(lijt|yijt ¼ c), γ c 1, γ min}] (Albert and
Chib 1993; Sorensen et al. 1995).

Variance of location effects (σ 2β If we give σ 2β χ 2 νβ ; Sβ ðshape and scaleÞ,
then
!
1 νβ Sβ
P σ β jELSE / P σ β P βj σ β ¼ νβ exp 2 P βj σ 2β
2 2 2
2 þ1 2σ β
σ 2β
!
1 ðβ β0 ÞT Σ1 ð β β Þ þ ν β S β
/ νβ þI exp 0 0
2 þ1
2σ 2β
σβ
2
This is the kernel of the scaled inverted χ 2 distribution; therefore, the full
conditional posterior is
h i
σ 2β j ELSE χ 2 ~ν β ¼ νβ þ I; ~
S β ¼ ðβ β0 ÞT Σ 1
0 ð β β 0 Þ þ ν β S β =ν β þ I
Appendix B: Derivation of Full Conditional Distributions

for Model BNBR
Full conditional for β∗
Y
I Y
J Y
nij
f ð β∗ jELSEÞ ¼ Pr Y ijt ¼ yijt jxiT , r , ωijt , b1j , b2ij f ð β∗ Þ
i¼1 j¼1 t¼1
!T
X 2
1 X 2
∗ ∗
/ exp κT X β þ κT Z h bh Xβ þ Zh bh
h¼1
2 h¼1
! !
X2
1
D ω X β∗ þ Zh bh ð β∗ β0 Þ Σ1 2 ∗
T
0 σ β ð β β0 Þ
h¼1
2
h
1
/exp β∗T Σ1 2 T ∗ 1 2
0 σ β þX Dω X β 2 Σ0 σ β β0 X Dω
T
2
P
2
Zh bh þXT κÞT β∗ Þ
h¼1
h
1 ∗ ~ T ~ 1 ∗ ~ i
/ exp β β0 Σ0 β β0 / N ~β 0 ; Σ
~0
2
1 P
2
~ 0 ¼ Σ1 σ 2 þ XT Dω X
Σ β0 ¼ Σ0 ðΣ1 2
0 σ β β0 X D ω Zh bh þ
T
where 0 β ,
h¼1
X κÞ.
T
Full conditional for ωijt

" #
ωijt ðxTi β∗ þ b1j þ b2ij Þ2
f ðωijt jELSEÞ / exp f ðωijt ; yijt þ r, 0Þ
2
" #
ωijt ðxTi β∗ þ b1j þ b2ij Þ2
/ exp f ðωijt ; yijt þ r, 0Þ
2
/ PGðyijt þ r, xTi β∗ þ b1j þ b2ij Þ
Full conditional for b1

Defining η1 ¼ X β∗ + Z2b2, the conditional distribution of b1 is given as

1
1 T

f ðb1 jELSEÞ / exp κ Z1 b1 Z1 b1 þ η Dω Z1 b1 þ η
T 1
f b1 jσ 2b1
2
h i
1 1
T
1 T
/ exp b1T σ 2 b1 G 1 þ Z T
D
1 ω 1Z u 2 Z 1 κ Z T
D
1 ω η b 1
2

1
/ exp ðb1 b1 ÞT F1 1 ðb1 b1 Þ Nðb1 , F1 Þ
2
1
where F1 ¼ σ 2 1
b1 G 1 þ Z 1 D ω Z 1
T
and ~
b 1 ¼ F1 Z1T κ Z1T Dω η1 .
Full conditional for σ 2bh
!
1 b T G1 bh þ νbh Sbh
f σ 2bh jELSE / exp h h
h 2 h þ1
νb þnb
2σ 2bh
σ 2bh

/ χ 2 ~ν b ¼ νbh þ nbh ; ~
S b ¼ bhT G1
h bh þ νbh Sbh =νbh þ nbh
with nb1 ¼ J and nb2 ¼ IJ.

Full conditional for σ 2β∗
!
1 ðβ∗ β0 ÞT Σ1 ∗
0 ðβ β0 Þ þ νβ∗ Sβ∗
f ðσ 2β∗ jELSEÞ / ν ∗ þI exp
β
2 þ1
2σ 2β∗
ðσ 2β∗ Þ

/ χ 2 νβ∗ ¼ νβ∗ þ I, Sβ ¼ ½ðβ∗ β0 ÞT Σ1 ∗
0 ðβ β0 Þ þ νβ∗ Sβ∗ =νβ∗ þ IÞ
Full conditional for r

To make the inference of r, we first place a gamma prior on it as r ~ G(a0, 1/b0).
Then we infer a latent count L for each Y NB(μ, r) conditional on Y and r. Since
L ~ Pois(r log(1 π)), by construction we can use the Gamma-Poisson conjugacy
to update r. Therefore,
Y
I Y nij
J Y
f ðrjELSEÞ / f ðr Þ f yijt jLijt f Lijt
i¼1 j¼1 t
Y
I Y
J Y
nij
L
/ r a0 1 expðrb0 Þ rlog 1 π ij ijt exp r log 1 π ij
i¼1 j¼1 t
P I P J Pnij I X
X J X
a0 þ Lijt 1 nij
/r i¼1 j¼1 t¼1 exp½ðb0 t¼1
logð1 π ij ÞrÞ
i¼1 j¼1
!
I X
X J X
nij 1
/ G a0 log 1 π ij ; PI PJ Pnij ð4:A:5Þ
i¼1 j¼1
t¼1 b0 þ i¼1 j¼1 t¼1 Lijt
According to Zhou et al. (2012), the conditional posterior distribution of Lijt is a

Chinese restaurant table (CRT) count random variable. That
yijt
is, Lijt ~ CRT(yijt, r) and
we can sample it as Lijt ¼ Σ l¼1 dl , where dl Bernoulli l1þr
r
:
References
Albert JH, Chib S (1993) Bayesian analysis of binary and polychotomous response data. J Am Stat
Assoc 88(422):669–679
Berridge DM, Crouchley R (2011) Multivariate generalized linear mixed models using R. CRC
Press, Boca Raton
Bartlett MS (1947) The use of transformations. Biometrics 3(1):39–52
Brier GW (1950) Verification of forecasts expressed in terms of probability. Mon Weather Rev
78:1–3
Burgue~no J, de los Campos GDL, Weigel K, Crossa J (2012) Genomic prediction of breeding
values when modeling genotype environment interaction using pedigree and dense molec-
ular markers. Crop Sci 52:707–719
Casellas J, Caja G, Ferret A, Piedrafita J (2007) Analysis of litter size and days to lambing in the
Ripollesa ewe. I. comparison of models with linear and threshold approaches . J Anim Sci
85:618–624
Cavanagh, C.R., Chao, S., Wang, S. et al. (2013). Genome-wide comparative diversity uncovers
multiple targets of selection for improvement in hexaploid wheat landraces and cultivars.
Proceedings of the National Academy of Sciences. 110(20):8057–8062
Crossa J, Pérez-Rodrı́guez P, de los Campos G, Mahuku G, Dreisigacker S, Magorokosho C
(2011) Genomic selection and prediction in plant breeding. Journal of Crop Improvement 25
(3):239–261
Czado C, Gneiting T, Held L (2009) Predictive model assessment for count data. Biometrics 65
(4):1254–1261
de los Campos, G., and Perez-Rodriguez, P. (2013). BGLR: Bayesian generalized linear regres-
sion. R package version. http://R-Forge.R-project.org/projects/bglr/
de Maturana EL, Gianola D, Rosa GJM, Weigel KA (2009) Predictive ability of models for
calving difficulty in US Holsteins. J Anim Breed Genet 126:179–188
Garthwaite PH, Kadane JB, O’Hagan A (2005) Statistical methods for eliciting probability
distributions. J Am Stat Assoc 100(470):680–701
Gelfand AE, Smith AF (1990) Sampling-based approaches to calculating marginal densities. J Am
Stat Assoc 85(410):398–409
Geyer CJ (1992) Practical Markov chain Monte Carlo. Stat Sci 7(4):473–483
Gianola D (1980) A method of sire evaluation for dichotomies. J of Anim Sci 51(6):1266–1271
Gianola D (1982) Theory and analysis of threshold characters. J Anim Sci 54(5):1079–1096
Gianola D, Foulley JL (1983) Sire evaluation for ordered categorical data with a threshold model.
Genet Sel Evol 15(2):1–23
Gianola D (2013) Priors in whole-genome regression: the Bayesian alphabet returns. Genetics
194:573–596
González-Camacho JM, de los Campos G, Pérez-Rodrı́guez P, Gianola D, Cairns JE, Mahuku G,
Crossa J (2012) Genome-enabled prediction of genetic values using radial basis function neural
networks. Theor Appl Genet 125(4):759–771
González-Recio O, Forni S (2011) Genome-wide prediction of discrete traits using Bayesian
regressions and machine learning. Genet Sel Evol 43:7
Hoerl AE, Kennard RW (1970) Ridge regression: biased estimation for nonorthogonal problems.
Technometrics 12(1):55–67
Kizilkaya K, Tait RG, Garrick DJ, Fernando RL, Reecy JM (2011) Whole genome analysis of
infectious bovine keratoconjunctivitis in Angus cattle using Bayesian threshold models. BMC
Proc 5:S22
Kizilkaya K, Fernando RL, Garrick DJ (2014) Reduction in accuracy of genomic prediction for
ordered categorical data compared to continuous observations. Genet Sel Evol 46(1):37. doi:
10.1186/1297-9686-46-37
Link WA, Eaton MJ (2012) On thinning of chains in MCMC. Methods Ecol Evol 3(1):112–115
MacEachern SN, Berliner LM (1994) Subsampling the Gibbs sampler. Am Stat 48(3):188–190
McCulloch CE, Searle SR (2001) Generalized, linear, and mixed models (1st ed.). Chichester:
Wiley. ISBN 0-471-19364-X.
Montesinos-López OA, Montesinos-López A, Pérez-Rodrı́guez P, de los Campos G, Eskridge
KM, Crossa J (2015a) Threshold models for genome-enabled prediction of ordinal categorical
traits in plant breeding. G3: Genes| Genomes| Genetics 5(1):291–300
Montesinos-López OA, Montesinos-López A, Crossa J, Burgue~ no J, Eskridge K (2015b)
Genomic-enabled prediction of ordinal data with Bayesian logistic ordinal regression. G3:
Genes|Genomes|Genetics 5(10):2113–2126. http://doi.org/10.1534/g3.115.021154
Montesinos-López OA, Montesinos-López A, Pérez-Rodrı́guez P, Eskridge K, He X, Juliana P,

Crossa J (2015c) Genomic prediction models for count data. J Agric Biol Environ Stat 20
(2):533–554
Montesinos-López A, Montesinos-López OA, Crossa J, Burgue~ no J, Eskridge K, Falconi-Castillo-
E, He X, Singh P, Cichy K (2016) Genomic Bayesian prediction model for count data with
genotype environment interaction. G3: Genes|Genomes|Genetics 6(5):1165–1177
Nelder JA, Wedderburn RWM (1972) Generalized linear models. J R Stat Soc A 135:370–384.
doi:10.2307/2344614
O’Hara RB, Kotze DJ (2010) Do not log-transform count data. Methods Ecol Evol 1(2):118–122
Park T, van Dyk DA (2009) Partially collapsed Gibbs samplers: illustrations and applications. J
Comput Graph Stat 18(2):283–305
Polson NG, Scott JG, Windle J (2013) Bayesian inference for logistic models using Pólya–gamma
latent variables. J Am Stat Assoc 108:1339–1349
Quenouille MH (1949) A relation between the logarithmic, Poisson, and negative binomial series.
Biometrics 5:162–164
Ramirez-Valverde R, Misztal I, Bertrand J, K. (2001) Comparison of threshold vs linear and
animal vs sire models for predicting direct and maternal genetic effects on calving difficulty in
beef cattle. J Anim Sci 79:333–338
R Core Team (2015) R: A language and environment for statistical computing. R Foundation for
Statistical Computing, Vienna, Austria. ISBN 3–900051–07-0, URL http://www.R-project.org/
Schurink A, Wolc A, Ducro B, Frankena K, Garrick D, Dekkers J, van Arendonk J (2012)
Genome-wide association study of insect bite hypersensitivity in two horse populations in
the Netherlands. Genet Sel Evol 44(1):31
Scott J, Pillow JW (2013) Fully Bayesian inference for neural models with negative-binomial
spiking. In: Pereira F, Burges CJC, Bottou L, Weinberger KQ (eds) Advances in neural
information processing systems 25. Cornell University, New York, pp 1898–1906
Sorensen DA, Andersen S, Gianola D, Korsgaard I (1995) Bayesian inference in threshold models
using Gibbs sampling. Genet Sel Evol 27(3):229–249
Stroup WW (2012) Generalized linear mixed models: modern concepts, methods and applications.
CRC Press, Boca Raton
Stroup WW (2015) Rethinking the analysis of non-normal data in plant and soil science. Agron J
107(2):811–827
Teerapabolarn K, Jaioun K (2014) An improved Poisson approximation for the negative binomial
distribution. Appl Math Sci 8(89):4441–4445
VanRaden PM (2008) Efficient methods to compute genomic predictions. J Dairy Sci 91
(11):4414–4423
Vazquez AI, Weigel KA, Gianola D, Bates DM, Perez-Cabal MA et al (2009) Poisson versus
threshold models for genetic analysis of clinical mastitis in US Holsteins. J Dairy Sci
92:5239–5247
Varona L, Misztal I, Bertrand J, K. (1999) Threshold-linear versus linear-linear analysis of birth
weight and calving ease using an animal model. Ii. Comparison of models. J Anim Sci
77:2003–2007
Villanueva B, Fernandez J, Garcia-Cortes LA, Varona L, Daetwyler HD, Toro MA (2011)
Accuracy of genome-wide evaluation for disease resistance in aquaculture breeding programs.
J Anim Sci 89:3433–3442
Wang CL, Ding XD, Wang JY, Liu JF, Fu WX, Zhang Z, Jin ZJ, Zhang Q (2013) Bayesian
methods for estimating GEBVs of threshold traits. Heredity 110(3):213–219
Wecker WE (1989) Assessing the accuracy of time series model forecasts of count observations. J
Bus Econ Stat 7(4):418–419
Wright S (1934) An analysis of variability in number of digits in an inbred strain of guinea pigs.
Yang W, Tempelman RJ (2012) A Bayesian antedependence model for whole genome prediction.
Genetics 190(4):1491–1501
Zucknick, M., and Richardson, S. (2014). MCMC algorithms for Bayesian variable selection in the
logistic regression model for large-scale genomic applications. Technical Report. http://arxiv.
org/abs/1402.2713.
Zhou M, Li L, Dunson D, Carin L (2012) Lognormal and gamma mixed negative binomial
regression. In machine learning: proceedings of the international conference on machine
learning. vol. 2012. p 1343. NIH Public Access.
Chapter 5
Genomic Selection for Small Grain
Improvement
Jessica E. Rutkoski, Jared Crain, Jesse Poland, and Mark E. Sorrells
Abbreviations
BB BayesB
BRR Bayesian ridge regression
CIMMYT International Maize and Wheat Improvement Center
DArT Diversity Array Technology
DHs Doubled haploids
DON Deoxynivalenol
ECs Environmental covariates
FHB Fusarium head blight
GBS Genotyping by sequencing
GEBV Genomic estimated breeding value
GS Genomic selection
GxE Genotype-by-environment interaction
h2 Heritability
HTP High-throughput phenotyping
LD Linkage disequilibrium
MAS Marker-assisted selection
MEs Mega-environments
J.E. Rutkoski
International Programs, College of Agriculture and Life Sciences, Cornell University, Ithaca,
NY 14853, USA
J. Crain • J. Poland
Department of Plant Pathology, Kansas State University, Manhattan, KS 66506, USA
M.E. Sorrells (*)
Department of Plant Breeding and Genetics, Cornell University, Ithaca, New York 14853,
USA
e-mail: mes12@cornell.edu

DOI 10.1007/978-3-319-63170-7_5
100 J.E. Rutkoski et al.
MET Multi-environment trials

MxE Marker-by-environment interaction
PS Phenotypic selection
QTL Quantitative trait loci, RR-BLUP, ridge-regression best linear
unbiased prediction
SNP Single nucleotide polymorphism
TPE Target population of environments
5.1 Introduction
Small grains include sorghum (Sorghum bicolor), wheat (Tritcum aestivum L.), oats
(Avena sativa L.), barley (Hordeum vulgare L.), rye (Secale cereale L.), rice (Oryza
sativa), millet (Pennisetum glaucum), and triticale (Triticosecale) but not the
pseudo-cereals such as buckwheat (Fagopyrum esculentum) and quinoa
(Chenopodium quinoa). Most of the small grain cereals are self-pollinated with
the exception of rye. Consequently, this chapter will focus on genomic selection
(GS) research conducted on wheat, oats, barley, and rice. GS in hybrid cereals will
be covered in Chapter 7.
Animal breeders initiated GS research, in part, because of the high cost of
phenotyping and the inability to replicate individual genotypes. Meuwissen et al.
(2001) conducted foundational work in GS development by simultaneously estimat-
ing all genetic marker effects. Their simulation results showed up to a 0.84 correla-
tion between estimated breeding values obtained through GS and the true breeding
value. Based on these results, they proposed that GS could have significant impact in
plant and animal breeding programs by using dense markers to predict performance
of individuals that did not have phenotypic records. The genetic gain per unit time
achieved by a breeding program can be summarized by the breeder’s equation:
irσA
G¼ ð5:1Þ
Y
where G is the gain per year, i is selection intensity, r is selection accuracy, σ A is the
square root of narrow-sense heritability, and Y is time in years to complete a cycle
of selection (Falconer and Mackay 1996). By combining GS with methods to
shorten the breeding cycle, significant gains should be achieved (Meuwissen et al.
2001), with gains proportional to the reduction in breeding cycle time. Plant
breeders lagged behind in the use of mixed models and pedigrees for predicting
breeding value because phenotyping was relatively less expensive and genotypes
using inbred lines could be replicated to increase heritability. However, that has
changed rapidly as plant breeders have begun to incorporate GS into their breeding
programs. Although it is not considered a small grain, among the earliest publica-
tions on GS in crops was a simulation study by Bernardo and Yu (2007) using maize
as an example. Using a population derived from a biparental cross of maize inbreds
5 Genomic Selection for Small Grain Improvement 101
as a training population, they found that the predicted increase in selection gain
from ridge-regression best linear unbiased prediction (RR-BLUP) was 18% greater
than that from marker-assisted recurrent selection for a highly heritable trait (h2 ¼
0.8) and 43% for a trait with low heritability (h2 ¼ 0.2). In 2009, Heffner et al.
(2009) published a review and interpretation paper on GS highlighting the potential
as well as the challenges of applying GS in an applied breeding program. They
highlighted the importance of estimating allele effects rather than genotype effects
using many unreplicated lines. They also noted that genotype-by-environment
interaction (GxE) is likely to be much more problematic for plant breeders than
for animal breeders. These were among the earliest publications on GS in plants,
and publications that followed shortly thereafter were also simulations. The section
that follows reviews the main body of literature on GS in small grains.
5.2 Overview of GS Research in Small Grains
At least 40 GS studies have been published in small grains to date (Table 5.1).
Twenty-nine of them were conducted in bread wheat, five in barley, two in oat and
rye, and one in durum wheat (Triticum turgidum L. spp. durum), perennial ryegrass
(Lolium perenne L.), and intermediate wheatgrass (Thinopyrum intermedium).
Across all studies Diversity Array Technology (DArT) was the most frequently
used marker platform followed by genotyping by sequencing (GBS) and single
nucleotide polymorphism (SNP). Taken together, these studies indicate that GS
could be successfully applied in cereals breeding to increase rates of genetic gain.
The first few GS studies in small grains were published between 2009 and 2011
(Crossa et al. 2010; Heffner et al. 2011a; de los Campos et al. 2009). Both de los
campos et al. (2009) and Cross et al. (Crossa et al. 2010) used data from the
International Maize and Wheat Improvement Center (CIMMYT) wheat breeding
program. A 13–42% increase in correlation between predicted values and observed
values was reported by de los Campos et al. (2009) using Bayesian LASSO
compared to prediction models using pedigree alone. Crossa et al. (2010) used
CIMMYT wheat breeding data from the international yield trials as well as
CIMMYT maize data to evaluate parametric and semi-parametric prediction
models based on pedigree and/or genomic relationship. Both de los Campos et al.
(2009) and Crossa et al. (2010) concluded that models that used genomic markers
were superior to those that used only pedigree relationships. Heffner et al. (Heffner
et al. 2011a) used data from a soft white wheat breeding program for multiple traits
to compare prediction models, GS to marker-assisted selection (MAS) and pheno-
typic selection (PS) accuracies, and to look at the impact of training population size
and number of markers on the GS accuracy. The authors concluded that based on
their results, GS could increase the rate of genetic gain per unit time and cost if
applied in a wheat breeding program. Many other studies of GS in cereals were
published thereafter, and here we highlight a few key studies.
Table 5.1 Genomic selection studies in small grains
Topics covered
Effect of
New training Other
Number of prediction population factors Coping Realized
Small grain Population markers and Prediction GS model model on affecting with GS genetic
a
Reference species Traits size platformb modelsc comparison development accuracy accuracy GxE vs. MAS gain
Arruda et al. Triticum Six Fusarium 273 5054 GBS RR-BLUP, x x
(2015) aestivum L. head blight LASSO, elastic
resistance traits net
Asoro et al. Avena sativa Beta-glucan 446 1005 DArT RR-BLUP, x x x
(2011) L. concentration, Bayes Cπ
days to heading,
groat percent,
plant height,
grain yield
Asoro et al. Avena sativa Beta-glucan 446 and 866 and G-BLUP x x
(2013) L. concentration 482 675 DArT
Burgue~no et al. Triticum Grain yield 599 1279 DArT G-BLUP, and x x x
(2012) aestivum L. G-BLUP model-
ing GxE
Crossa et al. Triticum Grain yield 599 1279 DArT RKHS, Bayesian x
(2010) aestivum L. LASSO,
G-BLUP
Crossa et al. Triticum Grain yield, 388 7594 SNP ME RR-BLUP x x x
(2016a), b) turgidum grain volume
L. spp. weight, 1000-
durum kernel weight,
days to heading
Daetwyler et al. Triticum Resistance to 206 5568 SNP G-BLUP, x
(2014) aestivum L. leaf rust, stem BayesR
rust, and yellow
rust based on
disease severity
scores
Dawson et al. Triticum Grain yield 622 34,843 GBS G-BLUP, x x
(2013) aestivum L. G-BLUP with
environment
cluster specific
variances,
G-BLUP with
environment
clusters covari-
ances modeled
as factor analytic
Endelman (2011) Triticum Grain yield 599 1279 DArT G-BLUP, BLUP x
aestivum L. with Gaussian
and exponential
kernels
Fè et al. (2015) Lolium Days to heading 1846 1,005,509 GBS G-BLUP x
perenne L.
He et al. (2016) Triticum Grain yield 2325 12,642 SNP RR-BLUP, x x
aestivum L. Bayes Cπ,
RKHS,
EG-BLUP
Heffner et al. Triticum Days to heading, 374 1158 DArT Multiple linear x x x
(2011a) aestivum L. plant height, regression,
lodging toler- RR-BLUP,
ance, preharvest BayesA,
sprouting toler- BayesB, Bayes
ance, grain Cπ
yield, flour
yield, and seven
other soft wheat
quality traits
Heffner et al. Triticum Preharvest 209 and 399 multiple plat- Stepwise regres- x x x
(2011b) aestivum L. sprouting toler- 174 forms and sion multiple
ance, flour yield, 574 DArT linear regres-
and seven other sion, RR-BLUP,
soft wheat qual- Bayes Cπ
ity traits
(continued)
Table 5.1 (continued)
Topics covered
Effect of
New training Other
a
Heslot et al. Triticum Grain yield 2437 1287 Factorial regres- x x x
(2014) aestivum L. sion and an
extension of fac-
torial regression
with soft rules
Heslot et al. Hordeum Grain yield 996 335 SNP Bayesian x x
(2013a) vulgare L. LASSO
Heslot et al. Triticum Grain yield, 365 1544 DArT and RR-BLUP x
(2013b) aestivum L. plant height, 38,412 GBS
heading date,
and preharvest
sprouting
tolerance
Heslot et al. Hordeum Grain yield, 761, 338 SNP, 2146 Bayes Cπ, x
(2012) vulgare beta-glucan con- 911, 599, SNP, 1279 DArT, Bayesian
L. and centration, head- 374, and 1158 DArT, and LASSO, BRR,
Triticum ing date, plant 551 319 SNP E-Bayes, NNET,
aestivum L. height, thousand RF, RKHS,
kernel weight RR-BLUP,
SVR, wBSR
Isidro et al. (2015) Triticum Grain yield, test 1127 38,893 GBS RR-BLUP x
aestivum L. weight, lodging
tolerance, days
to heading, and
plant height
Jarquı́n et al. Triticum Grain yield 139 2395 SNP A reaction norm x x x
(2014) aestivum L. model (exten-
sion of
G-BLUP)
Jiang et al. (2015) Triticum Fusarium head 372 782 SNP and SSR Multiple linear x x x
aestivum L. blight resistance combined regression,
index based on RR-BLUP;
disease severity RKHS; Bayes
and incidence Cπ
scores
Lado et al. (2016) Triticum Grain yield 1477 81,999 GBS G-BLUP and x x x x
aestivum L. G-BLUP model-
ing covariances
among
environments
Lopez-Cruz et al. Triticum Grain yield 693, 15,744, 15,744, BRR and BRR x x x
(2015) aestivum L. 670, and and 14,217 GBS modeling ME
807
Lorenz et al. Hordeum Fusarium head 691 3072 SNP RR-BLUP, x x x
(2012) vulgare L blight resistance Bayes Cπ,
based on sever- Bayesian
ity and DON LASSO
concentration
Mirdita et al. Triticum Fusarium head 2325 12,642 SNP RKHS, x x
(2015) aestivum L. blight resistance RR-BLUP
based on sever-
ity and Septoria
tritici blotch
resistance based
on severity
scores
Ornella et al. Triticum Resistance to 90–180 1400 DArT Bayesian x
(2012) aestivum L. yellow rust and including LASSO, RR,
stem rust, based non-polymorphic SVR
on severity markers
scores
Perez-Rodriguez Triticum Days to heading, 306 1717 DArT Bayesian x
et al. (2013) aestivum L. grain yield LASSO,
BayesA,
(continued)
Topics covered
Effect of
New training Other
a
BayesB, RKHS,
BRR, NNET
Poland et al. Triticum Grain yield, 254 41,371 GBS G-BLUP x
(2012) aestivum L. thousand kernel
weight, days to
heading
Rutkoski et al. Triticum Resistance to 365 4040 GBS Multiple linear x x x
(2014) aestivum L. stem rust regression,
G-BLUP,
Bayesian
LASSO, Bayes
Cπ
Rutkoski et al. Triticum Resistance to 626, 20,882, 20,882, G-BLUP and x
(2015a) aestivum L. stem rust 626, 1163, 18,653 and BRR
and 1150 18,653 GBS
Rutkoski et al. Triticum Resistance to 365 and 17,168 GBS G-BLUP x
(2015b) aestivum L. stem rust 503
Rutkoski et al. Triticum Five Fusarium 170 2402 DArT and RR-BLUP, x x x
(2012) aestivum L. head blight 38 SSR Bayesian
resistance traits LASSO, RKHS,
and days to RF, multiple lin-
heading ear regression
Sallam et al. Hordeum Fusarium head 647 1536 SSR RR-BLUP, x x
(2015) vulgare L blight resistance RKHS, Bayes
based on sever- Cπ
ity and DON
concentration,
grain yield, plant
height
Schmidt et al. Hordeum Twelve malting 65–424, 4095 and 4359 RR-BLUP x
(2016) vulgare L quality traits and 102 SNP
Schulthess et al. Secale Grain yield, pro- 201 and 394 and Multi-trait x x
(2016) cereale L. tein content 219 584 DArT RR-BLUP
Storlie and Triticum Grain yield 318 1279 DArT RR-BLUP, x
Charmet (2013) aestivum L. G-BLUP, BRR,
Bayesian
LASSO
Thavamanikumar Triticum Time to young 165 and 17,328 and BayesB, Bayes- x x
et al. (2015) aestivum L.microspore, 159 17,293 SNP ian LASSO,
spike grain RR-BLUP, PLS,
number and SPLS
Wang et al. Secale Grain yield, 220 and 1048 RR-BLUP and a x x x
(2014) cereale L. plant height, 220 regression
starch content, model using
total pentosan identified QTL
content
Ward et al. (2015) Triticum Days to heading, 151 603 DArT and RR-BLUP and x x
aestivum L. plant height, 655 metabolites differentially
grain yield, and penalized
21 quality traits regression
Zhang et al. Thinopyrum Heading score, 1126 3883 RR-BLUP, x x
(2016) intermedium plant height, G-BLUP with a
(host) head weight, Gaussian kernel,
Barkworth & threshability, BayesA,
D. R. Dewey seed weight, BayesB, Bayes
biomass Cπ, Bayesian
LASSO, BRR,
RKHS, and RF
Zhao et al. (2014) Triticum Days to heading, 1739 1280 SNP and RR-BLUP, x x x
aestivum L. plant height 3 function SNP Bayes Cπ,
markers RR-BLUP with
differentially
weighted predic-
tors, stepwise
(continued)
Topics covered
Effect of
New training Other
a
regression with
functional
markers
a
Fusarium head blight resistance is caused by Fusarium graminearum Schwabe, leaf rust is caused by Puccinia triticina, stem rust is caused by Puccinia graminis, yellow rust
is caused by Puccinia striiformis, DON deoxynivalenol, Septoria tritici blotch is caused by Zymoseptoria tritici
b
GBS genotyping by sequencing, DArT Diversity Array Technology, SNP single nucleotide polymorphism
c
RR-BLUP ridge-regression best liner unbiased prediction (Meuwissen et al. 2001; Whittaker et al. 2000), LASSO least absolute shrinkage and selection operator (Tibshirani
1996), elastic net (Zou and Hastie 2005), Bayesian LASSO (Park et al. 2008), Bayes Cπ (Habier et al. 2011), G-BLUP genomic best linear unbiased prediction (Habier et al.
2010), RKHS reproducing kernel Hilbert space regression (Gianola and Van Kaam 2008), ME marker-by-environment interaction, BayesR (Erbe et al. 2012), EG-BLUP
(Jiang and Reif 2015), BayesA (Meuwissen et al. 2001), BayesB (Meuwissen et al. 2001), BRR Bayesian ridge regression (Pérez et al. 2010), E-Bayes empirical Bayes
(Xu 2007), RF random forest regression (Breiman 2001), SVR support vector regression (Smola and Sch€ olkopf 2004), NNET neural network (Ripley 1996), wBSR weighted
Bayesian shrinkage regression (Hayashi and Iwata 2010), PLS partial least squares (Geladi and Kowalski 1986), SPLS sparse partial least squares (Chun and Keleş 2010)
In 2011, Asoro et al. published the first GS study in oat using data from an oat
breeding program for multiple traits to evaluate GS accuracies and to assess factors
affecting accuracy. This study was also the first GS study in small grains that
examined how characteristics of the training population affected GS accuracy.
The authors found that including older lines in the training set either increased or
had no impact on accuracy, suggesting that an appropriate model training popula-
tion can be constructed by accumulating breeding program data over time. The first
study in small grains that aimed to increase prediction accuracy by modeling
genotype-by-environment interaction was by Burgue~no et al. (Burgue~no et al.
2012). This study used the same data as in Crossa et al. (2010) and evaluated
whether modeling covariance between environments could improve prediction
accuracy. The results of this study indicated that modeling covariance between
environments could increase accuracy when predicting performance within specific
environments when the lines were observed in some environments but not others.
This highlights that phenotypes of the selection candidates observed in one envi-
ronment can be used to improve genomic prediction of selection candidates’
breeding values in other environments.
In 2012, the first study in small grains using GBS for genomic selection was
published by Poland et al. (2012). This study found that GBS led to greater GS
accuracies compared to current DArT markers. A later study by Heslot et al.
(2013b) followed up on this observation and found that GBS leads to higher GS
accuracies compared to DArT because GBS produces a larger number of
non-redundant, evenly distributed markers. Although many researchers are
concerned that the relatively large amount of missing data commonly observed in
GBS datasets may impede downstream analyses, Rutkoski et al. (2013) found that
the amount of missing data commonly observed in wheat GBS datasets has very
little impact on GS prediction accuracies.
The first report of realized gain from a GS experiment in small grains was
published by Asoro et al. in Asoro et al. 2013. Working in oats, the authors
compared GS, MAS, and phenotypic selection for beta-glucan concentration. In
both the GS and MAS schemes, phenotypic data on the selection candidates were
used in addition to marker data. The authors found that more superior individuals
originated from the populations developed using GS or MAS, demonstrating the
value of markers for improving selection. Rutkoski et al. (2015a) published the first
realized gain from GS experiment in wheat. This study compared GS with PS for
breeding for quantitative adult plant resistance to stem rust. Unlike in Asoro et al.
(2013), in the GS scheme, phenotypic data were not available on the selection
candidates prior to selection. The authors found that GS and PS lead to equal rates
of genetic gain per unit time, but GS led to a faster loss of genetic variance because
with GS two cycles were completed, while with PS only one cycle was completed.
The lack of improvement in genetic gain per unit time from GS over PS in this case
was due to the relatively low prediction accuracy in the first cycle of selection when
the model training population is not closely related to the selection candidates. The
University of Minnesota barley breeding program has implemented GS since 2010
for Fusarium head blight (FHB) resistance as well as for yield, winter hardiness,
and malting quality resulting in reduced time and labor for phenotyping (Bernardo
2016). In evaluating dynamic germplasm from the barley breeding program, Sallam
et al. (2015) found prediction accuracies that ranged from 0.03 to 0.99 for traits
including plant height, yield, FHB resistance, and deoxynivalenol (DON)
concentration.
5.3 Factors Affecting GS Prediction Accuracies
5.3.1 Theoretical Considerations of GS Accuracies
Utilizing foundational genetic theory, irrespective of species, we can predict GS

accuracy based on the heritability of the trait, the number of independent chromo-
some segments, and the number of individuals in the training population
(Daetwyler et al. 2010). For the simplest GS examples, this assumes (1) there is
perfect linkage between markers and quantitative trait loci (QTL), (2) the model
training and selection candidate individuals are sampled from the same population,
and (3) the trait of interest is conferred by a large number of additive loci. A large
body of work has extended these concepts to understand how GS performs in real-
world applications when germplasm and phenotypic data diverge from the ideal
conditions. For example, when markers are not in complete linkage disequilibrium
(LD) with QTL, increasing the number of markers so that more markers are in LD
with QTL leads to higher accuracies (Heffner et al. 2011b; Muir 2007). If model
training and selection candidates are sampled from different populations, the level
of relationship between the two populations is another factor affecting accuracy,
with increasing levels of relationship leading to higher accuracies (Pszczola et al.
2012). Lastly, when the trait of interest is conferred by few loci, the choice of
prediction model will affect accuracy.
5.3.2 Research to Increase GS Accuracy
Many of the research studies of GS in small grains have examined factors affecting
prediction accuracy. In small grains, at least 29 studies have looked at the effect of
prediction model on the GS accuracy, at least nine studies have examined how the
relationship between the training population and the selection candidate population
affects accuracy, and at least 16 studies have looked at other factors affecting
accuracy including training population size and number of markers.
5.3.3 Effect of GS Model
The vast majority of studies looking at the effect of prediction model on accuracy
report either no or only a small effect of the prediction model on the accuracy
(Heslot et al. 2012; Sallam et al. 2015). Among studies that have compared additive
and nonadditive GS models, Mirdita et al. (2015), He et al. (2016), Endelman et al.
(2011), Perez-Rodriguez et al. (2013), and Rutkoski et al. (2012) observed that
nonadditive GS models led to higher accuracies compared to additive models. In
contrast, Ornella et al. (2012) reported that in biparental populations, nonlinear
models led to lower accuracies compared to linear models.
5.3.4 Relationship of Training and Validation Population
Studies in small grains that have examined the effect of the relationship between the
individuals in the training and validation population on the GS accuracy have
observed that having a closer relationship between training and validation
populations leads to higher GS accuracies (Asoro et al. 2011; Lorenz et al. 2012;
Rutkoski et al. 2015b; Wang et al. 2014; Zhang et al. 2016). Interestingly, Rutkoski
et al. (2015b) found that including older model training data after updating the model
training population could lead to decreased prediction accuracy but only if the older
model training data had a low heritability. On the other hand, Lorenz et al. (2012)
found that including individuals from a different subpopulation in the model training
set neither increased nor decreased accuracy. Asoro et al. (2011) reported that
including older individuals in the model training population either slightly increased
or did not affect the accuracy. In a study by Zhang et al. (2016) about GS in
intermediate wheatgrass, the authors reported that accuracies increased with increas-
ing numbers of families and genotypes per family in the training set, but after
30 families and six genotypes per family, the increase in accuracy was very small.
The potential to select subsets of the model training population of a given size
that lead to the highest possible accuracy, referred to as training population
optimization, has been investigated for wheat by Isidro et al. (2015) and Rutkoski
et al. (2015b). Both studies found that training population optimization was better
than random sampling for selecting subsets from a population for model training.
This could be useful for selecting which set of lines to phenotype for prediction
model updating to ensure the highest accuracy given the resources available.
Overall, studies in small grains generally confirm that GS model training
populations should be related to the selection candidates and should be frequently
updated to maintain accuracy. While most envision updating the GS model with
phenotypic data generated on the lines selected in the breeding program because of
their superior breeding values, it may be wise to use training population optimiza-
tion to select a set of lines to phenotype specifically for GS model updating. This
would include lines that would ordinarily get discarded in the breeding program.
More research is required in this area to determine if the benefits of selecting lines
to phenotype specifically for model training outweigh the costs.
5.3.5 Effect of Number of Markers and Individuals
Studies in small grains that have examined the effect of the number of markers and
number of individuals in the training population on the GS accuracy have observed
similar trends where accuracy increases linearly with marker number and training
population size until a plateau is reached, and a plateau in accuracy is reached
sooner with marker number than with training population size (Arruda et al. 2015;
Heffner et al. 2011a, b; Lorenz et al. 2012). Heffner et al. (2011a) observed that for
predicting within biparental populations, increasing training population size from
24 to 96 leads to a linear increase in accuracy, while increasing marker number
beyond 256 did not improve prediction accuracy. For predicting within a population
of advanced breeding lines, Heffner et al. (2011b) found that accuracy increased
linearly with TP size from 96 to 288 and very gradually when the number of
markers was increased beyond 384. Lorenz et al. (2012) reported that in a popula-
tion of barley breeding lines from different breeding programs, accuracy reached a
plateau at a population size of 200 and that marker number could be decreased to
384 without losing accuracy. Other work in barley by Sallam et al. (2015) found
that DON concentration level predictions plateaued at a TP of 75, while grain yield
prediction did not plateau based on the TP size, suggesting the TP size may be trait
specific. The occurrence of plateaus in accuracy at relatively low numbers of
markers and population sizes is a reflection of the low rate of LD decay with
physical distance, or in other words, a low number of independent chromosome
segments in the small grain breeding populations are used for GS studies.
In a population consisting of advanced wheat breeding lines from breeding
programs across the Midwestern and Eastern United States, Arruda et al. (2015)
found that for most traits, a plateau in accuracy was reached when the training
population was larger than 192 lines, and, depending on the trait, a plateau in
accuracy occurred when the number of markers was greater than 1500 or 3000. In
a breeding scenario where older breeding lines are used to predict newer breeding
lines, larger training population sizes and more markers may be required compared
to what cross validation studies would indicate because generations of recombina-
tion have taken place leading to faster rates of LD decay among the combined model
training-selection candidate population. This was observed in a study by Rutkoski
et al. (2015b) where the authors examined the increase in accuracy with increasing
training population size for two training sets, one distantly related and one closely
related to the selection candidates. They found that accuracy increased linearly with
training population size in the more distantly related training set, while accuracy
reached a plateau at 292 individuals with the more closely related training set.
Research on training population size and number of markers will ultimately need
to be conducted using forward validation within individual breeding programs.
5.4 Breeding Methods for Variety Development in Cereals
While any implementation of GS involves identifying high-performing individuals

based on a genomic prediction model, the methods that different breeding programs
could use may look quite different depending upon the budget and resources available
to the breeding program, the relative cost of genotyping vs. phenotyping, the gener-
ation time that can be achieved, and the heritability of the traits of interest. An
important feature of GS is that it is complementary to MAS. Marker-assisted
selection is most effective for simply inherited, high heritability traits, whereas GS
is relatively more effective for low heritability traits conferred by many QTL. Both
methods could be readily incorporated into a molecular breeding program and used in
concert to enable breeders to select for simply inherited traits with MAS and for
quantitative traits using GS. Applications such as spiked GBS (Rife et al. 2015) which
combine whole-genome profiling along with known marker assays could allow
breeding programs to efficiently and affordably integrate both GS and MAS.
5.4.1 Timing of GS Application Within Breeding Programs
Implementation of GS in cereal crops can be imposed in early or late generations

(Fig. 5.1). In early generation implementation, GS is applied to the selection
candidates directly after crossing prior to any selfing or after one generation of
selfing. Implementing GS in this way leads to a greater reduction in the breeding
cycle duration because the two or more growing seasons that would normally be
needed for selfing are eliminated (Heffner et al. 2010; Hickey et al. 2014). Selected
individuals are cycled back into the crossing block as parents. This rapid cycling
(Fig. 5.1) program allows individuals selected based on their genomic estimated
breeding value (GEBV) to be planted, cross- or self-pollinated, and harvested two
or more times a year for many cereal species. This would ordinarily not be possible
because many important traits must be evaluated on fixed lines using relatively
large quantities of seed. One consideration for implementing GS in this way is that
model updating will need to come from inbred lines derived from the selection
candidates in earlier cycles and potentially several cycles of GS selection could be
conducted for every single cycle of training population updating that is completed.
Once early generation lines are selected as parents in the rapid cycling program,
they can begin the inbreeding phase where culling based on MAS and PS can be
applied at any generation until the F4 or F5 generations as preferred. At that stage,
the candidates should be whole-genome genotyped and phenotyped to train the
model to predict GEBVs in the breeding population. The genome-wide marker data
and phenotypic data can also be fit in a model that captures nonadditive genetic
effects to improve selection accuracy on the lines per se for promotion as varieties.
Phenotyping and genotyping all selection candidates have been shown to be more
favorable for improving rates of genetic gain compared to only phenotyping a
Fig. 5.1 Integration of GS in a pure line breeding program. In the rapid cycling phase, GS is used
to enhance gain per unit time. In the inbreeding phase, MAS and PS can be imposed until the F4 or
F5 generation, and then whole-genome genotyping is used to select individuals that enter the
training population or are recycled in the crossing program. Each phase is conducted simulta-
neously, and the GS models are updated annually
subset of the genotyped selection candidates (Endelman et al. 2014). GS can also be
applied only among lines, without a rapid cycle program. Compared to a rapid
cycling approach, applying GS only among lines would enable higher selection
accuracies but would not reduce the breeding cycle duration as dramatically.
Different GS breeding schemes should be evaluated for each specific situation to
understand the trade-offs in cycle time and selection accuracy and to optimize gain
from GS. Deterministic and stochastic simulations can be useful tools for this
purpose.
Most of the GS research in cereal crops has been conducted using inbred lines. In
a deterministic simulation study by Heffner et al. (2010), GS among inbred lines
was found to increase gain from selection per unit time twofold compared to MAS
among inbred lines for known QTL. Examining how to optimize preliminary yield
trials, Endelman et al. (2014) found that a 5% increase over phenotypic selection
could be achieved by using GS in inbred lines.
Other authors have reported evaluations of various GS schemes (Longin et al.
2015). The study by Longin et al. (2015) evaluated GS using GS alone, GS followed
by one or two rounds of PS, and the comparison to PS only. At an estimated GS
accuracy of 0.3, the authors found that GS followed by one round of PS produced
the highest genetic gain; however, the rankings of methods were dependent upon
GS accuracy. For example, if GS accuracy could exceed 0.65 using GS only with no
PS provided the highest genetic gain, but under more realistic conditions
(GS accuracy 0.3), combinations of GS and PS were more productive. The use of
doubled haploid lines would preclude the use of MAS and PS during the inbreeding
phase of line development (Fig. 5.1).
5.4.2 Genomic Selection for Germplasm Improvement

Through Introgression of Alleles
While the previous GS strategies have focused mainly on reducing the length of the
breeding cycle or increasing the selection accuracy to increase the rate of genetic
gains, GS could also be used for introgression of exotic alleles into elite germplasm.
Crossa et al. (2016a) evaluated GS in landraces populations including over 2000
Iranian and 8000 Mexican lines. They found good GS prediction accuracies even
with GxE and population structure. They proposed that GS could be used to predict
genotype performance of all genotyped accessions within a germplasm collection
and then phenotyping could be conducted on the most promising lines, followed by
introgressing the selected exotic alleles into elite germplasm. Bernardo (2009)
simulated the effect of introgressing exotic alleles into adapted maize germplasm.
His results showed that GS could be used to rapidly incorporate exotic alleles into
elite germplasm. Additionally, this work provided some guidance on the number of
cycles of GS that could be used with exotic alleles and where to apply GS in the
introgression program. He found that the best starting material for GS was an F2
cross between exotic and adapted germplasm rather than a BC1 or BC2. In general
the rate of genetic progress declined after 7–8 cycles of GS, but assuming three
cycles of GS could be completed per year resulted in an equivalent time frame of
two rounds of PS (2 years per cycle). The 7–8 cycles of GS resulted in 1.25–2.4
times the rate of gain compared to the two cycles of PS depending on the number of
favorable alleles in the exotic germplasm and a trait heritability of 0.8. Another
simulation study in maize suggested that GS should be applied in exotic-by-exotic
crosses and used to increase the frequency of favorable alleles (Gorjanc et al. 2016).
This was because using GS with exotic-by-elite crosses quickly resulted in
reconstructing elite material. In practice, applying GS among exotic-by-exotic
crosses could be problematic due to poor adaptation. For example, if the exotic
individuals are photoperiod sensitive in the environments of interest, it will not be
possible to gather meaningful grain yield data for prediction modeling. Importantly,
Bernardo (2009) found that the quality (accuracy) of the phenotypic data and the
quantity (number) of phenotyped individuals were crucial for ensuring success of
GS in introgressing exotic germplasm. Thus, training population size and compo-
sition needs to be carefully considered. Bernardo (2009) suggested that if trait
heritability was low (h2¼0.2), then more field testing (replication) across multiple
environments should be used to increase the entry-mean heritability (accuracy).
Along with replication to improve the entry-mean heritability, Bernardo (2009)
suggested that training populations should be larger, 288 compared to 144 in

adapted-by-adapted crosses. Additionally, to achieve sufficiently high prediction
accuracy, GS for introgression of exotic germplasm would need to be within
biparental populations, which have slower rates of LD decay and fewer segregating
chromosome segments compared to populations derived from multiple families.
5.4.3 Combining Genomics and Phenomics for Increased

Precision
As reported by Endelman et al. (2014) and Lorenz (2013), phenotypic data on the
selection candidates per se can be used in GS models to increase selection accuracy
and gain from selection. With decreasing genotyping costs, phenotypes are quickly
becoming the most valuable asset to breeding programs. Field-based phenomics or
high-throughput phenotyping (HTP) is an active area of research that is working to
provide image and sensor data for traits that are correlated with the phenotypes of
interest (Cobb et al. 2013; White et al. 2012) which could be useful for prediction
modeling. Using a variety of proximal sensors, researchers have mapped QTL for
biomass growth in triticale (Busemeyer et al. 2013) as well as assessing differences
in crop response to well-watered and drought conditions in cotton (Andrade-
Sanchez et al. 2014). The ability to generate large volumes of data quickly and at
multiple time points possibly before grain yield testing has led to efforts to combine
phenotypic data within the GS model. There are several methods by which both GS
and HTP could be integrated into a breeding program. Low-cost HTP could be used
to evaluate lines in early generations to both provide data to train GS models and
make screening decisions when thousands of lines need to be evaluated (Araus and
Cairns 2014). Along with providing information about the crop, HTP could allow
breeding programs to increase population size of material evaluated while increas-
ing selection intensity resulting in higher genetic gain (Crain and Reynolds 2016).
Additionally, HTP and GS combinations could also be used for evaluating more
advanced lines. Using spectral indices and canopy temperature, Rutkoski et al.
(2016) and Crain et al. (unpublished) reported higher GS prediction accuracies by
including HTP data in the GS model. Rutkoski et al. (2016) found up to a 70%
increase in prediction accuracy for grain yield by including HTP traits of canopy
temperature and vegetation indices. By including canopy temperature and spectral
reflectance (Crain et al. unpublished) found an average of a 12% increase in GS
model accuracy compared to GS models utilizing only marker information. By
using multiple traits (phenotypes) from HTP data, the goal is to enhance model
prediction accuracy. Previous work with multiple traits has shown that prediction
accuracy of a low heritability trait can be greatly increased when a second or
multiple traits that have higher heritability are added to the model (Jia and Jannink
2012). They found that with a low heritability trait (h2 ¼ 0.1), prediction accuracy
went from 0.49 to 0.64 assuming a genetic correlation of 0.1. As the genetic
correlation increased, the prediction accuracy further increased. Along with

increasing prediction accuracy, they also found that prediction accuracies could
be increased when there was missing information. While these are some of the
earliest efforts to incorporate phenotypic data, HTP is offering many possibilities to
further increase the genetic gains from GS. The utilization of HTP along with GS
has the potential to provide scientists with rich datasets that adequately reflect real-
world conditions. For example, Montesinos-López et al. (2016) utilized the same
data presented by Rutkoski et al. (2016) to develop multi-trait, multi-environment
models. Employing these types of complex models that account for the genotypic
relationship between multiple traits and environmental factors (multiple environ-
ments) should allow breeders to get a more complete picture of how genetics are
expressed in different conditions allowing for more accurate selection decisions to
be made.
5.4.4 Additional Breeding Program Considerations
Along with increasing the accuracy of phenotyping, the design of the breeding
program can be modified to take full advantage of GS. With whole-genome
genotyping, testing more lines at different locations, rather than all lines at one
location or fewer lines at all locations, Endelman et al. (2014) found that prediction
accuracies were increased. This potentially surprising result comes from the fact
that in GS, the alleles are under selection rather than an individual genotype. Thus,
in field trials the goal is replication of alleles, but the replication of alleles is not
limited to specific genotypes (Lorenz et al. 2011).
The adept use of data and training populations has allowed GS to predict new
genotype performance in multiple environments as well as hybrid performance.
Lado et al. (2016) found that using related genotype information at multiple
locations could be used to predict new genotype performance with high (0.5)
accuracy. Zhao et al. (2015) used GS to predict performance of hybrid lines to
develop heterotic groups. While this application was for hybrid breeding, the ability
to predict hybrid (cross) performance without testing the phenotype could be
applicable for the pure line breeder as well. Both of these examples along with
others (e.g., Endelman et al. 2014; Hickey et al. 2014) highlight the importance of
training population design (size and quality) and genotyping (number of markers)
in making accurate predictions.
GS can be incorporated into a breeding program in several different facets.
Breeders could use GS to provide more information about the selections they are
making or to make all selection decisions. In practice, GS will probably be applied
somewhere in between these two extremes whether it be to increase favorable
alleles, rapidly cycle germplasm, or introgress exotic alleles. The success that GS
has will be dependent on the resourcefulness of the researcher and his/her ability to
fully utilize molecular, phenotypic, and environmental information in an efficient
way based on careful assessment of different breeding schemes.
5.5 GS for Analyzing and Predicting GxE
Plant breeders are concerned about GxE because it amplifies the phenotypic
variation without contributing to the additive genetic variation, thereby reducing
heritability across environments. The challenges of GxE have been with plant
breeders since the beginning. However, whole-genome genotyping and GS present
some opportunities to the modern plant breeder that are already changing our
strategies for dealing with GxE, for both genotype selection and multiple environ-
ment trial evaluation methods.
5.5.1 Target Population of Environments
Fundamental to the concept of GxE is the definition and sampling of a target

population of environments (TPE). The TPE represents the biotic and abiotic
factors that released varieties are likely to encounter during their production.
Because genotypes respond differently to environmental factors, in multi-
environment trials (MET), varieties perform differently resulting in relative differ-
ences in performance as well as rank changes. Rank changes or crossover interac-
tions complicate selection because one variety is not the best for all environments.
The challenge is to adequately sample environments, especially over years where
climate can be more variable. Generally, the composition of the TPE is unknown.
Years with insufficient precipitation may be more common than those with ade-
quate precipitation, and a streak of dry years could bias the evaluation of experi-
mental lines. One approach to managing environmental variability is to classify
environments based on historical data and then weight the trials by their expected
frequency of occurrence in the TPE (Podlich et al. 1999). This approach effectively
adjusts for the negative effects of unrepresentative environments on selection
resulting in increased gain from selection.
5.5.2 Application of GS Models to GxE
GxE was recognized as a major issue for applying GS in plants (Crossa et al. 2010;
Heffner et al. 2009). Crossa (2012) reviewed GxE and marker-by-environment
interaction studies and discussed models for assessing marker effects, QTL, and
marker effect by environment interactions and for studying the pattern of
covariability of marker effects across environments. He made the case for evaluating
different models from different areas of statistical research to better understand
genetic effects and their interaction with environment. Studies that have evaluated
GS predictions that model genotype-by-environment interaction report an increase
in accuracy from modeling the interaction rather than ignoring it (Burgue~no et al.
2012; Heslot et al. 2014; Jarquı́n et al. 2014; Lado et al. 2016; Lopez-Cruz et al.
2015). Within GS accounting for GxE remains an active area of research with
numerous methods proposed for utilizing GxE including marker-by-environment
interactions (Heslot et al. 2013a), using genetic correlations between TPE
(Burgue~ no et al. 2012), and using environmental covariates (Heslot et al. 2014).
5.5.3 GxE by Modeling Marker Replication and Interaction
By design, plant breeding MET data are unbalanced, thus limiting the kinds of
analyses that can be performed on the data because all genotypes are not
represented in all environments. The transforming principle first employed by
Heslot et al. (2013a) is that even though all genotypes are not represented in all
environments, all marker effects are represented in all environments. This allows
one to measure the relative similarities among environments based on marker
effects as well as similarities based on prediction of marker effects. This approach
is especially useful in cases where a factor analytic model fails convergence.
Similarities based on marker effects can be determined using Euclidean distances
and visualized using cluster analysis (Fig. 5.2, Heslot et al. 2013a). Using multi-
environment yield trial data from a commercial barley breeding program, the
authors found that outlier environments were readily identified. Additionally, the
breeder field notes corroborated the results; however, for this dataset, grouping
environments based on similarity of marker effects did not increase prediction
accuracy. Likewise, prediction of marker effects in other environments can be
calculated with a simple correlation analysis and visualized in a heat map. Although
the patterns were not as clear as the marker effects, grouping environments based on
average reciprocal prediction accuracies increased prediction accuracy for yield
across environments (Heslot et al. 2013a).
In a second experiment designed to optimize the composition of the training
population, Heslot et al. (2013a) used the average predictive ability of each
environment for predicting performance of lines in the other environments in the
same dataset. The environments were then ranked from least predictive to most
predictive, and starting with the least predictive, one environment was removed at a
time, and then the model was retrained on the remaining environments to determine
if prediction accuracy improved (Fig. 5.3, Heslot et al. 2013a). The environments
that were removed were placed in an unpredictive set, and the prediction accuracy
of that set was calculated. When the prediction accuracy of the predictive set
dropped and/or the unpredictive set increased, the remaining environments were
considered to be the optimal set, and, in this study, accuracy was increased from
0.54 to 0.61. Out of the 58 environments, 18 unpredictive environments were
removed. Interestingly, some outlier environments were included, and only one
barley line was excluded in the optimal set of environments.
Fig. 5.2 Heat map showing the similarity of environments based on Euclidian distances computed
using marker effects. Environment comparisons with red shading are more dissimilar, and those
environments with blue shading are more similar (Fig. 5.3 from Heslot et al. 2013a)
Another benefit from whole-genome genotyping is that by modeling MxE, we

can begin to understand what genetic regions have main (stable) effects and which
ones interact with the environment. Lopez-Cruz (2015) modeled MxE using regres-
sion of wheat phenotypes on markers or using covariance structures (a genomic best
linear unbiased prediction-type model) to estimate main (stable) effects and
environment-specific (interaction) effects. Environments that were correlated
exhibited low MxE, and those that were not correlated showed high MxE. Modeling
MxE in general improved prediction accuracy over models that did not take MxE
into account. The MxE model limits the ability to interpret patterns of GxE that are
not positive leading the authors to conclude that the model is best suited for the joint
analysis of positively correlated environments (Lopez-Cruz et al. 2015).
Model development to represent GxE is a research area that is continually
advancing. To further extend research on MxE, Crossa et al. (2016b) explored the
use of priors that produce shrinkage and variable selection including Bayesian ridge
regression (BRR) and BayesB (BB) in durum wheat. They evaluated the genomic
prediction accuracy of MxE models within and across environments. The MxE
model minimized the model residual variance and improved data-fitting gain for
Fig. 5.3 Optimization of the training population. The blue dots are cross-validated accuracies for
the selected training population (predictive set), and red triangles are prediction accuracies for the
environments removed from the training population (unpredictive set). Green squares are the
prediction accuracies for a validation set observed in 2011 (Fig. 5 from Heslot et al. 2013a)
more simply inherited traits compared to more complex traits such as grain yield
and test weight. The MxE model identified markers for the major genes for heading
date including Ppd-A1, Ppd-B1, and TaFT-A on chromosomes 2A, 2B, and 7A, and
their effects were stable across environments. For grain yield, several additional
chromosome regions with large marker effects were identified in all chromosome
groups. Another example of modeling GxE was given by Cuevas et al. (2016) in
which nonlinear Gaussian kernels were used to model MxE. Because this model
allowed for small, complex MxE interactions, they were able to capture up to 60%
greater predictions compared to models using a single environment.
5.5.4 GxE by Treating Environments as Multiple Traits
Genotype-by-environment interactions can be analyzed by treating different envi-

ronments as multiple traits and considering variety performance in different envi-
ronments as correlated traits (Falconer and Mackay 1996) or, in the case of strong
GxE, a lack of correlation among environments. Consequently, genetic correlations

among environments in the TPE can be used to increase the prediction accuracy
across environments in the target region (Burgue~ no et al. 2012). In one of the earlier
papers involving GxE, Burgue~ no et al. (2011) compared linear mixed models and
factor analytic models for their predictive ability. When GxE was important,
modeling GxE using the factor analytic model improved prediction accuracy;
otherwise when GxE was not significant, most models gave relatively high predic-
tion accuracies. Burgue~ no et al. (2012) examined wheat MET using GS multi-
environment (multi-trait) models and evaluated their predictive accuracy with and
without pedigree and marker information. In their cross validation, they predicted
either the performance of untested genotypes or the performance of genotypes that
had been evaluated in only some environments. Models that included both markers
and pedigrees were superior to those that included either alone. Additionally,
prediction accuracies were higher for predicting the performance of genotypes in
untested environments than for predicting untested genotypes. They concluded that
prediction accuracy could be improved using multi-environment GS models.
Modeling GxE when genotypic and environmental data are highly dimensional
can present computational problems. More recent papers have focused on marker-
by-environment interaction (MxE) effects. Jarquin et al. (2014) proposed a variance
components approach where they used covariance functions to model high-
dimensional interactions between markers and environmental covariates (ECs) for
wheat and maize. In principle, it should be possible to model GxE by regressing
phenotypes on markers and ECs and partitioning the GxE. In the reaction norm
model, genetic and environmental gradients are described using a linear regression
on genetic markers and on ECs. They used 68 ECs related to different crop
developmental stages and compared interactions with main effects. Prediction
models that included the interaction terms were 17–34% more accurate than models
based only on main effects.
Lado et al. (2016) also used the correlations among environments to design sets
of environments having low GxE to try and better predict genotype performance in
untested environments. They used mixed models to generate the variance–covari-
ance matrix across environments in a large, highly unbalanced, historical dataset
from a wheat breeding program to obtain predictions within or across different sets
of environments. They grouped environments into three mega-environments (MEs)
based on a genotype-by-GxE biplot. The best predictions were within years across
locations or within MEs for a given year or location. They concluded that borrow-
ing information from environments using a variance–covariance matrix was useful
for predicting new genotypes prior to phenotyping. Cuevas et al. (2017) presented
results using a Bayesian genomic kernel model to account for the correlation
between environments. This model accounting for GxE always showed superiority
to models that only assessed one environment.
5.5.5 Dissecting GxE Using Environmental Covariates

and Crop Models
As described above, GxE can be taken into account using multiplicative mixed
models such as the factor analytic structure to model the covariance between
environments responsible for GxE. However, those approaches have numerical
limitations because of the highly unbalanced nature of multi-environment plant
breeding datasets. Also, because they are based on observed covariance among
environments, they are only explanatory of past performance rather than predictive
of future performance. Another interesting approach to better understand and
predict GxE involved the integration of ECs into the genomic selection framework
to predict GxE deviations for unobserved environments (Heslot et al. 2014).
Including environmental covariates in the analysis presented some of the same
issues encountered using GS methods such as a high number of covariates, each
explaining a small amount of the total variance while being highly correlated with
each other.
Heslot et al. (2014) modeled genome-wide markers and their differential
response to the environment to better understand the genetic architecture of GxE.
Using more than 2000 winter wheat lines grown in 44 environments over 6 years in
France, daily weather data (AGRI4CAST), and a wheat crop model known as
SirusQuality (Martre et al. 2006), they first synchronized the developmental stages
of the crop with the climatic conditions during those stages. Stress covariates
(climatic variables at a specific developmental stage) were derived by developmen-
tal stage by using knowledge about the sensitivity of specific growth stages to
abiotic stresses. The stress covariates were then used as independent variables in
statistical genetic models for effect estimation and prediction. The factorial regres-
sion model was extended to the genomic selection context, and for each marker,
they fit a main effect and a sensitivity to each of the stress covariates. A machine-
learning algorithm was used to capture the interactions between markers and stress
covariates as well as nonlinear effects. Genotype performance was predicted as a
main effect plus a GxE deviation.
To deal with the high dimension of n markers by n covariate predictors, they
assessed the variance of marker effects across environments and eliminated those
explaining little or no variation. The photoperiod sensitivity gene Ppd-D1 had the
highest variance but alone did not capture a significant part of the GxE variance.
The optimal model based on cross validation used 250 markers plus the nonlinear
soft rule fit component. The most important stress covariate was the sum of the
average daily temperature between meiosis and flowering. The second most impor-
tant was drought in the early spring measured by “total number of dry days to
350 degree days” and the “sum of precipitation and evapotranspiration potential.”
Heat stresses before flowering and during early grain fill were also important
covariates.
A factor analytic model predicted a GxE response for any genotype in any
environment, even if an environment had no phenotypic data for that genotype.
Euclidean distances could be calculated for all environments based on the predicted
level of genetic correlation between environments, and a cluster analysis can be
used to reveal the structure of the TPE. By including the GxE component, they were
able to increase prediction accuracy for genotype performance in unobserved
environments by 11.1% on average and the variability in prediction accuracy
decreased by 10.8%. In contrast to the approaches that used covariances among
environments to improve prediction accuracy, the use of carefully selected stress
covariates allowed for prediction in unobserved environments rather than a retro-
spective view of GxE. This approach provides important information to the breeder
because it offers a mechanism to leverage agronomy and physiology knowledge,
reduce dimensionality and nonlinearity, use existing breeding data, and interpret
results to identify specific environmental stresses.
Although their research involved maize, it is important to consider the applica-
tion of the crop growth model and whole-genome prediction proposed by Technow
et al. (2015) and empirically evaluated by Cooper et al. (2016). Cooper et al. (2016)
were able to predict drought tolerance of a set of doubled haploids (DHs) (from the
same cross) in maize hybrids using five measures of crop growth in a model that
incorporated whole-genome prediction and an algorithm based on approximate
Bayesian computation. As expected, prediction accuracies were high when
predicting entries in the same environment (0.53–0.82) but generally low when
predicting test DH entries in a new environment (0.22–0.38). If this approach is able
to deliver a means for predicting genotype performance in the environments and
management practices of the TPE, the breeder could use predicted performance of
genotypes for important environment types of the TPE instead of using predictions
based only on performance across environments.
5.6 Cost Benefit Analysis of GS
The breeder’s equation (Eq. 1) provides the information needed to optimize selec-
tion strategies for different GS applications. While manipulation of any of the
variables in the formula can drive the rate of genetic gain, the majority of GS
work has focused on shortening the length of time per cycle as the increase in
genetic gain would increase proportionally. Any strategy that increases genetic gain
could be used in a breeding program; however, only strategies that allow GS to
surpass the rate of gain achieved by PS or where GS is sufficiently cheaper than PS
will find applications in breeding programs.
Heffner et al. (2010) addressed this question by comparing the cost and genetic
gain per unit time for conventional MAS and GS. Their results indicated that GS
could achieve greater genetic gain than MAS on a per-year basis, even when GEBV
accuracies are low. They predicted that given a prediction accuracy of 0.5, the
expected annual gain from GS would exceed that of MAS by up to threefold for a
high-intensity maize breeding program and up to twofold for a low-intensity winter
wheat breeding program. The advantage realized by GS was almost entirely due to
the shorter breeding cycle. The case for decreasing breeding cycle time has also
been shown for introgressing wild alleles where Bernardo (2009) showed that the
rate of genetic gain using 7–8 cycles of GS was higher than two cycles of
phenotypic test-cross selection in maize.
Genomic selection may also find use where the trait of interest is challenging to
measure. For example, FHB susceptibility and mycotoxin DON levels require
laborious and expensive phenotyping. Using GS to predict FHB susceptibility and
DON levels, Lorenz et al. (2012) found high prediction accuracy 0.72 and 0.68 for
FHB and DON, respectively. These accuracies were equal to phenotypic selection
and were estimated to cost only 25% as much as PS. Rutkoski et al. (2012) also
reported high prediction accuracies (>0.62) for DON when markers were used in
combination with phenotypes.
Early work in wheat quality by Heffner et al. (2011a) found that low marker
density (256 markers) and small training population sizes (96 genotypes) in bipa-
rental crosses resulted in 0.66 ratio between the GS prediction and PS. Assuming
that two GS cycles could be completed per year, they estimated that GS could
provide more gains than PS for all nine of the wheat quality traits studied at 1/3rd
the cost of PS. Further work in wheat by Battenfield et al. (2016) demonstrated that
application of GS for predicting milling and baking quality traits in wheat had the
potential to substantially outperform PS. The prediction accuracies for complex and
expensive phenotypes in the CIMMYT bread wheat breeding program such as
mixing time and loaf volume in the quality lab were moderate ranging from 0.32
(grain hardness) to 0.62 (mixing time). However, even with these moderate levels
of prediction accuracy, the lower cost and higher throughput of genotyping relative
to phenotyping for milling and baking gave substantial advantage to GS. Based on
the current implementation of ten times more samples being genotyped in the
breeding program than the capacity to phenotype for quality, they calculated a
1.4–2.7 higher rate of genetic gain for GS over PS for the quality traits.
Heslot et al. (2015) made a case for consideration of the problem of optimal
resource allocation to obtain maximum genetic gains. One example of optimizing
resources (Endelman et al. 2014) evaluated the optimal design of the preliminary
yield trial. Using biparental barley and maize populations, they found that up to a
5% increase in genetic gain could be achieved if genotyping was 25% of the cost of
one yield plot unit assuming a breeding program with 250 yield plot units per
family. While GS prediction accuracy shows positive genetic gain, more cost
benefit studies are needed (Heslot et al. 2015).
5.7 Summary
Breeding programs are dynamic entities, and consideration of the cost at each stage
is required to optimize gain. For example, there will be questions about what
germplasm to use, correlations among traits, trade-offs between family size and
number of families, balance between phenotypic and GS or MAS at a constant
budget, relationship between the training or mapping population, and the breeding
germplasm. Some of the efficiencies that can be realized with GS and whole-
genome genotyping include increased gain by genotyping individuals that are
also phenotyped, more efficient experimental designs (sparse testing), reduced
nursery sizes, and selection for costly traits or traits that are not expressed in each
season. While researchers have worked to provide answers to many of these
questions (e.g., Endelman et al. 2014; Heslot et al. 2015; Hickey et al. 2014),
there are still many uncharted courses for how GS could play within individual
breeding programs that are driven by different goals, environments, and policies.
GS has considerable potential for improving quantitative traits, and new approaches
for implementation of GS will continue to evolve in applied breeding programs.
References
Andrade-Sanchez P, Gore MA, Heun JT, Thorp KR, Carmo-Silva AE, French AN et al (2014)
Development and evaluation of a field-based high-throughput phenotyping platform. Funct
Plant Biol 41(1):68–79
Araus JL, Cairns JE (2014) Field high-throughput phenotyping: the new crop breeding frontier.
Trends Plant Sci 19(1):52–61
Arruda MP, Brown PJ, Lipka AE, Krill AM, Thurber C, Kolb FL (2015) Genomic selection for
predicting fusarium head blight resistance in a wheat breeding program. Plant Genome 8
(3):1–12
Asoro FG, Newell MA, Beavis WD, Scott MP, Jannink J-L (2011) Accuracy and training
population design for genomic selection on quantitative traits in elite North American oats.
Plant Genome J 4(2):132
Asoro FG, Newell MA, Beavis WD, Scott MP, Tinker NA, Jannink JL (2013) Genomic, marker-
assisted, and pedigree-BLUP selection methods for β-glucan concentration in elite oat. Crop
Sci 53(5):1894–1906
Battenfield SD, Guzmán C, Gaynor RC, Singh RP, Pe~ na RJ, Dreisigacker S et al (2016) Genomic
selection for processing and end-use quality traits in the CIMMYT spring bread wheat breeding
program. Plant Genome 9(2):1–12
Bernardo R (2009) Genomewide selection for rapid introgression of exotic germplasm in maize.
Crop Sci 49(2):419
Bernardo R (2016) Bandwagons I, too, have known. Theor. Appl. Genet. Springer, Berlin
Heidelberg. 129(12):2323–2332.
Bernardo R, Yu J (2007) Prospects for genomewide selection for quantitative traits in maize. Crop
Sci 47(3):1082–1090
Breiman L (2001) Random forests. Mach Learn 45(1):5–32
Burgue~no J, Crossa J, Cotes JM, Vicente FS, Das B (2011) Prediction assessment of linear mixed
models for multienvironment trials. Crop Sci 51(3):944–954
Burgue~no J, de los Campos G, Weigel K, Crossa J (2012) Genomic prediction of breeding values
when modeling genotype x environment interaction using pedigree and dense molecular
markers. Crop Sci 52(2):707–719
Busemeyer L, Ruckelshausen A, M€ oller K, Melchinger AE, Alheit KV, Maurer HP et al (2013)
Precision phenotyping of biomass accumulation in triticale reveals temporal genetic patterns of
regulation. Sci Rep 3:2442
Chun H, Keleş S (2010) Sparse partial least squares regression for simultaneous dimension
reduction and variable selection. J R Stat Soc Ser B Stat Methodol 72(1):3–25
Cobb JN, DeClerck G, Greenberg A, Clark R, McCouch S (2013) Next-generation phenotyping:

requirements and strategies for enhancing our understanding of genotype-phenotype relation-
ships and its relevance to crop improvement. Theor Appl Genet 867–887.
Cooper M, Technow F, Messina C, Gho C, Totir LR (2016) Use of crop growth models with
whole-genome prediction: application to a maize multienvironment trial. Crop Sci 56:1–16
Crain JL, Reynolds MP, Poland JA (2016) Utilizing high-throughput phenotypic data for improved
phenotypic selection of stress adaptive traits in wheat. Crop Sci
Crossa J (2012) From genotype x environment interaction to gene x environment interaction. Curr
Genomics 13(3):225–244
Crossa J, Jarquı́n D, Franco J, Pérez-Rodrı́guez P, Burgue~ no J, Saint-Pierre C et al (2016a)
Genomic prediction of gene bank wheat landraces. G3 Genes|Genomes|Genetics 6
(7):1819–1834
Crossa J, de los Campos G, Maccaferri M, Tuberosa R, Burgue~ no J, Pérez-Rodrı́guez P (2016b)
Extending the marker X environment interaction model for genomic-enabled prediction and
genome-wide association analysis in durum wheat. Crop Sci 56(5):2193–2209
Crossa J, De Los CG, Pérez P, Gianola D, Burgue~ no J, Araus JL et al (2010) Prediction of genetic
values of quantitative traits in plant breeding using pedigree and molecular markers. Genetics
186(2):713–724
Cuevas J, Crossa J, Montesinos-Lopez O, Burgueno J, Perez-Rodriguez P, de los Campos G (2017)
Bayesian genomic prediction with genotype environment interaction kernel models. G3
Genes|Genomes|Genetics 7(1):41–53
Cuevas J, Crossa J, Soberanis V, Pérez-Elizalde S, Pérez-Rodrı́guez P, de los Campos G et al
(2016) Genomic prediction of genotype x environment interaction kernel regression models.
Plant Genome 9(3):1–20
Daetwyler HD, Bansal UK, Bariana HS, Hayden MJ, Hayes BJ (2014) Genomic prediction for rust
resistance in diverse wheat landraces. Theor Appl Genet 127(8):1795–1803
Daetwyler HD, Pong-Wong R, Villanueva B, Woolliams JA (2010) The impact of genetic
architecture on genome-wide evaluation methods. Genetics 185(3):1021–1031
Dawson J, Endelman J, Heslot N (2013) The use of unbalanced historical data for genomic
selection in an international wheat breeding program. F Crop Res 154:12–22
de los Campos G, Naya H, Gianola D, Crossa J, Legarra A, Manfredi E et al (2009) Predicting
quantitative traits with regression models for dense molecular markers and pedigree. Genetics
182(1):375–385
Endelman JB (2011) Ridge regression and other kernels for genomic selection with R package
rrBLUP. Plant Genome 4(3):250–255
Endelman JB, Atlin GN, Beyene Y, Semagn K, Zhang X, Sorrells ME et al (2014) Optimal design
of preliminary yield trials with genome-wide markers. Crop Sci 54(1):48–59
Erbe M, Hayes BJ, Matukumalli LK, Goswami S, Bowman PJ, Reich CM et al (2012) Improving
accuracy of genomic predictions within and between dairy cattle breeds with imputed high-
density single nucleotide polymorphism panels. J Dairy Sci 95(7):4114–4129
Falconer DS, Mackay TFC (1996) Quantitative Genetics, 4th edn. Pearson, New York
Fè D, Cericola F, Byrne S, Lenk I, Ashraf BH, Pedersen MG et al (2015) Genomic dissection and
prediction of heading date in perennial ryegrass. BMC Genomics 16(1):921
Geladi P, Kowalski BR (1986) Partial least-squares regression: a tutorial. Anal Chim Acta
185:1–17
Gianola D, Van Kaam JBCHM (2008) Reproducing kernel Hilbert spaces regression methods for
genomic assisted prediction of quantitative traits. Genetics 178(4):2289–2303
Gorjanc G, Jenko J, Hearne SJ, Hickey JM (2016) Initiating maize pre-breeding programs using
genomic selection to harness polygenic variation from landrace populations. BMC Genomics
17(1):30
Habier D, Fernando RL, Kizilkaya K, Garrick DJ (2011) Extension of the bayesian alphabet for
genomic selection. BMC Bioinformatics 12:186
Habier D, Tetens J, Seefried F-R, Lichtner P, Thaller G (2010) The impact of genetic relationship
information on genomic breeding values in German Holstein cattle. Genet Sel Evol 42(1):5
Hayashi T, Iwata H (2010) EM algorithm for Bayesian estimation of genomic breeding values.
BMC Genet 11:3
He S, Schulthess AW, Mirdita V, Zhao Y, Korzun V, Bothe R et al (2016) Genomic selection in a
commercial winter wheat population. Theor Appl Genet 129(3):641–651
Heffner EL, Jannink JL, Iwata H, Souza E, Sorrells ME (2011a) Genomic selection accuracy for
grain quality traits in biparental wheat populations. Crop Sci 51(6):2597–2606
Heffner EL, Jannink JL, Sorrells ME (2011b) Genomic selection accuracy using multifamily
prediction models in a wheat breeding program. Plant Genome 4(1):65–75
Heffner EL, Lorenz AJ, Jannink JL, Sorrells ME (2010) Plant breeding with genomic selection:
gain per unit time and cost. Crop Sci 50(5):1681–1690
Heffner EL, Sorrells ME, Jannink J-L (2009) Genomic selection for crop improvement. Crop Sci
49(1):1–12
Heslot N, Akdemir D, Sorrells ME, Jannink JL (2014) Integrating environmental covariates and
crop modeling into the genomic selection framework to predict genotype by environment
interactions. Theor Appl Genet 127(2):463–480
Heslot N, Jannink J-L, Sorrells ME (2015) Perspectives for genomic selection applications and
research in plants. Crop Sci 55(1):1–12
Heslot N, Jannink JL, Sorrells ME (2013a) Using genomic prediction to characterize environments
and optimize prediction accuracy in applied breeding data. Crop Sci 53(3):921–933
Heslot N, Rutkoski J, Poland J, Jannink JL, Sorrells ME (2013b) Impact of marker ascertainment
bias on genomic selection accuracy and estimates of genetic diversity. PLoS One 8(9)
Heslot N, Yang H-PP, Sorrells MEMEME, Jannink J-LL (2012) Genomic selection in plant
breeding: a comparison of models. Crop Sci 52(1):146
Hickey JM, Dreisigacker S, Crossa J, Hearne S, Babu R, Prasanna BM et al (2014) Evaluation of
genomic selection training population designs and genotyping strategies in plant breeding
programs using simulation. Crop Sci 54(4):1476–1488
Isidro J, Jannink J-L, Akdemir D, Poland J, Heslot N, Sorrells ME (2015) Training set optimization
under population structure in genomic selection. Theor Appl Genet 128:145–158
Jarquı́n D, Crossa J, Lacaze X, Du Cheyron P, Daucourt J, Lorgeou J et al (2014) A reaction norm
model for genomic selection using high-dimensional genomic and environmental data. Theor
Appl Genet 127(3):595–607
Jia Y, Jannink JL (2012) Multiple-trait genomic selection methods increase genetic value predic-
tion accuracy. Genetics 192(4):1513–1522
Jiang Y, Reif JC (2015) Modeling epistasis in genomic selection. Genetics 201(2):759–768
Jiang Y, Zhao Y, Rodemann B, Plieske J, Kollers S, Korzun V et al (2015) Potential and limits to
unravel the genetic architecture and predict the variation of Fusarium head blight resistance in
European winter wheat (Triticum aestivum L.) Heredity (Edinb) 114:318–326
Lado B, González Barrios P, Quincke M, Silva P, Gutiérrez L (2016) Modeling genotype x
environment interaction for genomic selection with unbalanced data from a wheat breeding
program. Crop Sci 56(April):1–15
Genet Springer Berlin Heidelberg 128(7):1297–1306
Lopez-Cruz M, Crossa J, Bonnett D, Dreisigacker S, Poland J, Jannink J-LJ et al (2015) Increased
prediction accuracy in wheat breeding trials using a marker x environment interaction genomic
selection model. G3 Genes|Genomes|Genetics 5(4):569–582
Lorenz AJ (2013 Mar) Resource allocation for maximizing prediction accuracy and genetic gain of
genomic selection in plant breeding: a simulation experiment. G3 Genes|Genomes|Genetics 3
(3):481–491
Lorenz AJ, Chao S, Asoro FG, Heffner EL, Hayashi T, Iwata H et al (2011) Genomic selection in
plant breeding. Adv Agron 110:77–122
Lorenz AJ, Smith KP, Jannink JL (2012) Potential and optimization of genomic selection for
Fusarium head blight resistance in six-row barley. Crop Sci 52(4):1609–1621
Martre P, Jamieson PD, Semenov MA, Zyskowski RF, Porter JR, Triboi E (2006) Modelling
protein content and composition in relation to crop nitrogen dynamics for wheat. Eur J Agron
25(2):138–154
Meuwissen THE, Hayes BJ, Goddard ME (2001) Prediction of total genetic value using genome-
wide dense marker maps. Genetics 157(4):1819–1829
Mirdita V, He S, Zhao Y, Korzun V, Bothe R, Ebmeyer E et al (2015) Potential and limits of whole
genome prediction of resistance to Fusarium head blight and Septoria tritici blotch in a vast
Central European elite winter wheat population. Theor Appl Genet 128(12):2471–2481
Montesinos-López OA, Montesinos-López A, Crossa J, Toledo F, Pérez-Hernández O, Eskridge
KM et al (2016) A genomic bayesian multi-trait and multi-environment model. G3 Genes|
Genomes|Genetics 6(9):2725–2744
Muir WM (2007) Comparison of genomic and traditional BLUP-estimated breeding value accu-
racy and selection response under alternative trait and genomic parameters. J Anim Breed
Genet 124(6):342–355
Ornella L, Singh S, Perez P, Burgue~ no J, Singh R, Tapia E et al (2012) Genomic prediction of
genetic values for resistance to wheat rusts. Plant Genome J 5(3):136–148
Park T, Casella G, Ark TP, Asella GC (2008) The Bayesian lasso. J Am Stat Assoc 103
(482):681–686
Perez-Rodriguez P, Gianola D, Gonzalez-Camacho JM, Crossa J, Manes Y, Dreisigacker S (2013)
Comparison between linear and non-parametric regression models for genome-enabled pre-
diction in wheat. G3 Genes|Genomes|Genetics 2(12):1595–1605
Pérez P, de los Campos G, Crossa J, Gianola D (2010) Genomic-enabled prediction based on
molecular markers and pedigree using the Bayesian linear regression package in R. Plant
Genome 3(2):106–116
Podlich DW, Cooper M, Basford KE (1999) Computer simulation of a selection strategy to
accommodate genotype-environment interactions in a wheat recurrent selection programme.
Plant Breed 118(1):17–28
Poland J, Endelman J, Dawson J, Rutkoski J, Wu S, Manes Y et al (2012) Genomic selection in
wheat breeding using genotyping-by-sequencing. Plant Genome J 5(3):103
Pszczola M, Strabel T, Mulder HA, Calus MPL (2012) Reliability of direct genomic values for
animals with different relationships within and to the reference population. J Dairy Sci 95
(1):389–400
Rife TW, Wu S, Bowden R, Poland JA (2015) Spiked GBS: a unified, open platform for single
marker genotyping and whole-genome profiling. BMC Genomics 16(1):1–7
Ripley BD (1996) Pattern recognition and neural networks. Cambridge University Press. Cam-
bridge, UK
Rutkoski J, Benson J, Jia Y, Brown-Guedira G, Jannink J-L, Sorrells M (2012) Evaluation of
genomic prediction methods for Fusarium head blight resistance in wheat. Plant Genome J 5
(2):51
Rutkoski J, Poland J, Mondal S, Autrique E, González Párez L, Crossa JJ et al (2016) Canopy
temperature and vegetation indices from high-throughput phenotyping improve accuracy of
pedigree and genomic selection for grain yield in wheat. G3 Genes|Genomes|Genetics 6
(9):2799–2808
Rutkoski J, Singh RP, Huerta-Espino J, Bhavani S, Poland J, Jannink JL et al (2015a) Genetic gain
from phenotypic and genomic selection for quantitative resistance to stem rust of wheat. Plant
Genome 8(2):1–10
Rutkoski J, Singh RP, Huerta-Espino J, Bhavani S, Poland J, Jannink JL et al (2015b) Efficient use
of historical data for genomic selection: a case study of stem rust resistance in wheat. Plant
Genome 8(1):1–10
Rutkoski JE, Poland JA, Singh RP, Huerta-espino J, Barbier H, Rouse MN et al (2014) Genomic
selection for quantitative adult plant stem rust resistance in wheat. Plant Genome 7(3):1–10
Rutkoski JE, Poland J, Jannink JL, Sorrells ME, Breeding P, York N (2013) Imputation of
unordered markers and the impact on genomic selection accuracy. G3 Genes|Genomes|Genet-
ics 3(3):427–439
Sallam AH, Endelman JB, Jannink J-L, Smith KP (2015) Assessing genomic selection prediction
accuracy in a dynamic barley breeding population. Plant Genome 8(1):1–15
Schmidt M, Kollers S, Maasberg-Prelle A, Großer J, Schinkel B, Tomerius A et al (2016)
Prediction of malting quality traits in barley based on genome-wide marker data to assess
the potential of genomic selection. Theor Appl Genet 129(2):203–213
Schulthess AW, Wang Y, Miedaner T, Wilde P, Reif JC, Zhao Y (2016) Multiple-trait- and
selection indices-genomic predictions for grain yield and protein content in rye for feeding
purposes. Theor Appl Genet 129(2):273–287
Smola AJ, Sch€olkopf B (2004) A tutorial on support vector regression. Stat Comput 14:199–222
Storlie E, Charmet G (2013) Genomic selection accuracy using historical data generated in a wheat
breeding program. Plant Genome 6(1):1–9
Technow F, Messina CD, Totir LR, Cooper M (2015) Integrating crop growth models with whole
genome prediction through approximate Bayesian computation. PLoS One 10(6):e0130855
Thavamanikumar S, Dolferus R, Thumma BR (2015) Comparison of genomic selection models to
predict flowering time and spike grain number in two hexaploid wheat doubled haploid
populations. G3 Genes|Genomes|Genetics 5(October):1991–1998
Tibshirani R (1996) Regression selection and shrinkage via the lasso. J R Stat Soc B 128:267–288
Wang Y, Mette MF, Miedaner T, Gottwald M, Wilde P, Reif JC et al (2014) The accuracy of
prediction of genomic selection in elite hybrid rye populations surpasses the accuracy of
marker-assisted selection and is equally augmented by multiple field evaluation locations
and test years. BMC Genomics 15(1):556
Ward J, Rakszegi M, Bedő Z, Shewry PR, Mackay I (2015) Differentially penalized regression to
predict agronomic traits from metabolites and markers in wheat. BMC Genet 16(1):1–7
White J, Andrade-Sanchez P, Gore MA, Bronson KF, Coffelt TA, Conley MM et al (2012) Field-
based phenomics for plant genetics research. F Crop Res 133:101–112
Genet Res 75(2):249–252
Xu S (2007) An empirical Bayes method for estimating epistatic effects of quantitative trait loci.
Biometrics 63(2):513–521
Zhang X, Sallam A, Gao L, Kantarski T, Poland J, DeHaan LR et al (2016) Establishment and
optimization of genomic selection to accelerate the domestication and improvement of inter-
mediate wheatgrass. Plant Genome 9(1):1–18
Zhao Y, Li Z, Liu G, Jiang Y, Maurer HP, Würschum T et al (2015) Genome-based establishment
of a high-yielding heterotic pattern for hybrid wheat breeding. Proc Natl Acad Sci U S A 112
(51):15624–15629
Zhao Y, Mette MF, Gowda M, Longin CFH, Reif JC (2014) Bridging the gap between marker-
assisted and genomic selection of heading time and plant height in hybrid wheat. Heredity
(Edinb) 112(6):638–645
Zou H, Hastie T (2005) Regularization and variable selection via the elastic-net. J R Stat Soc 67
(2):301–320
Chapter 6
Current Status and Prospects of Genomic
Selection in Legumes
Ankit Jain, Manish Roorkiwal, Manish K. Pandey, and Rajeev K. Varshney
6.1 Introduction
Availability of proper nutrition is of extreme importance as malnutrition at an early

age may lead to reduced physical and mental development and limits the capacity to
learn. UN World Food Program has reported that more than 900 million people in
the world do not get nutritious food to eat. Global population has been growing at a
fast pace, and feeding the ever increasing population with nutritious food is
becoming more difficult day by day. This will continue until there is significant
genetic gain by increasing crop productivity with enhanced nutrition. Although
significant efforts have been focussing on enhancing the crop production to feed the
world, still there are famines occurring in several parts of the world (http://www.
latimes.com/world/africa/la-fg-southsudan-famine-20170220-story.html). Consid-
ering this alarming situation, the United Nations and other affiliated organizations
have a challenge to eradicate hunger and malnutrition to ensure food and nutrition
security by responding to nutritional needs, addressing emerging threats and meet-
ing the zero hunger challenge. To overcome this devastating situation of malnutri-
tion, legumes are expected to play significant role, and there is a dire need to
enhance the productivity of these legumes.
Legumes have been cultivated since early civilizations and have been the major
source of nutrition for humans and animals (Power 1987; Graham and Vance 2003;
Varshney et al. 2013a; Rubiales and Mikic 2015; Pandey et al. 2016). Legumes
have been recognized as most valuable food to meet the dietary requirements of
undernourished or underserved global populations (Rebello et al. 2014). Research
has shown that replacement of energy dense foods with legumes offers various
A. Jain • M. Roorkiwal (*) • M.K. Pandey • R.K. Varshney

International Crops Research Institute for the Semi-Arid Tropics (ICRISAT), Patancheru,
Telangana, India
e-mail: m.roorkiwal@cgiar.org

DOI 10.1007/978-3-319-63170-7_6
132 A. Jain et al.
health benefits (Tarawali and Ogunbile 1995). In addition, legumes have the ability
to fix atmospheric nitrogen, which is vital for improving the soil nutritional profile,
thereby reducing the requirement for nitrogen fertilizers enabling legumes more
suited for crop rotation programs.
Legumes are among the important crop commodities and have high demand
being a major supplement of protein, but the productivity is low compared with the
increasing demand resulting from several biotic (Rubiales and Mikic 2015) and
abiotic stresses (Araújo et al. 2015). The productivity trends for these legumes in
the last five decades suggest very little improvement leading to low productivity in
most of the legumes compared with cereal crops (FAOSTAT 2014). Nevertheless,
several efforts made in these years identified the genetic variations for various traits
of interest in these legumes to enhance the crop productivity. So far, limited success
could be achieved with the application of conventional breeding approaches for
enhancing the crop productivity by overcoming key constraints. It is time to adopt
modern and new technologies for enhancing the rate of genetic gain, so that
improved varieties can be developed faster and more precisely equipped with
essential traits to face the climate and other stress factors.
A paradigm shift is required in approaches and breeding methodologies to
develop superior varieties for the future. In this context, deployment of genomics
tools and technologies has shown great potential in understanding the complex
genetics and breeding problems. It has been realized that genomics-assisted breed-
ing (GAB), with integration of conventional breeding is the key to overcome
conventional breeding limitations (Varshney et al. 2013a). Further in the case of
legumes, a journey from a status of orphan crops with a dearth of genomic resources
a decade ago, to current well-enriched genomic resource crop status, opened the
possibility of deployment of GAB for these crops. Additionally, recent advent of
the next-generation sequencing (NGS) technologies had brought down the sequenc-
ing and genotyping cost significantly. As a result, draft genomes have become
available for several legume crops including model legumes, i.e., Medicago
truncatula (Young et al. 2011), Lotus japonicus (Sato et al. 2008) and crops such
as Glycine max (Soybean) (Schmutz et al. 2010), Cajanus cajan (Pigeonpea)
(Varshney et al. 2012), Cicer arietinum (Chickpea) (Varshney et al. 2013b; Jain
et al. 2013); Lupinus angustifolius (Lupin) (Yang et al. 2013), Vigna radiata (Mung
bean) (Kang et al. 2014) and Arachis duranensis and A. ipaensis (progenitors of
cultivated groundnut) (Bertioli et al. 2016; Chen et al. 2016). Genome sequencing
efforts followed by large scale re-sequencing efforts in each crop led to availability
of millions of structural variations leading to availability of large numbers of
genetic markers (see Varshney et al. 2013a; Bohra et al. 2014; Pandey et al. 2016).
Availability of large scale genome-wide genetic markers led to establishment of
several high-throughput genotyping platforms, offering precise, rapid and cost-
effective solutions to genotyping of large populations. For instance, informative
single nucleotide polymorphisms (SNPs) with high genome density are being
chosen and used to design assays/platforms for legumes such as in Vigna
unguiculata (Egbadzor et al. 2014; Huynh et al. 2013; Lucas et al. 2013, Mu~noz-
Amatriaı́n et al. 2016), Pisum sativum (Deulvot et al. 2010; Bordat et al. 2011;
Tayeh et al. 2015), Lens culinaris (Sharpe et al. 2013; Kaur et al. 2014a), Vicia faba
6 Current Status and Prospects of Genomic Selection in Legumes 133
(Kaur et al. 2014b), soybean (Lee et al. 2015; Wang et al. 2016), chickpea (Gujaria
et al. 2011; Hiremath et al. 2011; Roorkiwal et al. 2014), pigeonpea (Saxena et al.
2012) and groundnut (Pandey et al. 2017). Other alternative SNP detection systems
like competitive allele-specific PCR (KASPar) (Cottage et al. 2012; Hiremath et al.
2012; Kumar et al. 2012; Saxena et al. 2012; Xu et al. 2012; Fedoruk 2013; Khera
et al. 2013; Sharpe et al. 2013), custom-designed Illumina VeraCode assay
(Deulvot et al. 2010; Roorkiwal et al. 2013, Duarte et al. 2014) have also been
employed for various applications. The development and deployment of different
genotyping platforms provide cost effective and precise genotyping solution to
many legume crops leading to enhanced rate of progress in legume genomics.
NGS-based genotyping by sequencing (GBS) allows simultaneous marker discov-
ery as well as genotyping of the populations even in the absence of a reference
genome (Davey et al. 2011). Among legumes, the GBS approach has been success-
fully used in lentil (Ates et al. 2016) and chickpea (Deokar et al. 2014; Jaganathan
et al. 2015; Verma et al. 2015) for genome-wide SNP discovery and genetic
mapping. Further, whole genome re-sequencing (WGRS) and restriction site-
associated DNA (RAD) sequencing approaches have also been used to capture
the variations in the genome and to understand diversity prevailing in the germ-
plasm (see Varshney et al. 2013b).
GAB aims at to accelerate crop improvement by establishing and exploiting the
relationships between genotype and phenotype. Of the three GAB approaches,
marker-assisted backcrossing (MABC), marker-assisted recurrent selection
(MARS) and genomic selection (GS), MABC has been deployed in most of the
crops and proved to be an effective approach for development of improved varieties
and lines in many legume crop plants (see Pandey et al. 2016). MABC uses markers
linked to agronomical important traits and mainly aims at introgression of a limited
number of alleles from one genetic background (donor) to other (recipient) (Hos-
pital 2005). Further, the improved varieties developed as a result of MABC contain
one or a few alleles at major gene/QTLs from the donor genotype, keeping intact
the rest of the genome from recurrent parent (see Varshney et al. 2013a). For
instance, one “QTL-hotspot” region having QTLs for several drought tolerance-
related root traits was introgressed into JG11, a desi chickpea cultivar from the
drought tolerant line ICC4958 (Varshney et al. 2013c). Similarly introgression lines
developed using MABC for fusarium wilt (FW) and ascochyta blight
(AB) resistance in the background of C214 have shown enhanced resistance for
FW and AB (Varshney et al. 2014). In the case of groundnut, MABC has been
exploited to introgress major QTLs for leaf rust resistance from GPBD 4, a leaf rust
resistant cultivar into ICGV 91114, JL 24 and TAG 24 cultivars (Varshney et al.
2014). MABC along with MAS was further deployed in enhancing the oil quality
by increasing oleic acid in three different groundnut varieties, viz. ICGV 06110,
ICGV 06142 and ICGV 06420 (Janila et al. 2016). In the case of pea, Aphanomyces
root rot resistance QTLs (Lavaud et al. 2015) and frost tolerance QTLs (Hascoët
et al. 2014) were introgressed using MABC into different agronomically important
genetic backgrounds. Likewise in soybean, MABC was deployed successfully to
improve resistance to a defoliating insect (Zhu et al. 2007), bacterial leaf pustule
134 A. Jain et al.
resistance (Kim et al. 2008) and to reducing a kunitz trypsin inhibitor (Kumar
et al. 2015).
In order to address the limitations of MABC approach for improving multiple
complex traits, MARS has been proposed for combining major and minor QTLs in
several crops. In the case of MARS, the de novo QTL identification is carried out in
a breeding population derived from the crosses of superior varieties followed by
crossing genotypes with superior alleles for pyramiding targeted QTLs into one or
more genetic backgrounds (Bernardo and Charcosset 2006). However, the MARS
approach was not effective for increasing yield in chickpea (Pandey et al. 2016).
MARS was suggested a method for improvement of drought tolerance in ground-
nut, however more than 100 main and epistatic effect QTLs were reported
because handling these small effect QTLs through MABC was not possible
(Gautami et al. 2012).
GS utilizes phenotypic as well as genome-wide marker data to predict the
genomic-estimated breeding values (GEBV) for selecting the superior lines. In
brief, two populations, training population and testing population (sometimes, it
is part of training population, hence known as validation set as well) are used.
Training population is the one with comprehensive phenotypic data under different
environmental conditions, that is, different locations/seasons/treatments. Genome-
wide genotypic and phenotypic data for the training population are used to train
different statistical GS models. The training population can be subdivided into five
to ten groups, and then, cross validation is used to evaluate the GS models and
prediction accuracy. Trained models, are used to calculate GEBV of a testing or
selection candidate population that has been genotyped but not phenotyped. The
predicted GEBVs are used to select superior lines from the population. One of the
advantages associated with GS is that it reduces the selection cycle length by
eliminating the phenotyping that is required for multiple rounds of selection
hence reducing time and cost, leading to genetic gain.
Genomic prediction is a key to success in GS breeding, and it depends on high-
throughput and high-density genotyping along with accurate, multilocation
phenotyping data. Availability of ample genomic resources and affordable high-
density and high-throughput genotyping in several legumes will facilitate deploy-
ment of GS in legumes. This chapter briefly describes the critical factors determin-
ing the success of genomic selection and summarises the ongoing efforts to deploy
genomic selection in legumes and further the existing possibilities by integrating
available genomic resources to harness the full potential of modern breeding
approaches.
6.2 Critical Factors in Deployment of Genomic Selection
High-precision prediction accuracies are the most critical point that determines the
success of any GS breeding program. Multiple simulation and empirical studies
involving estimation of prediction accuracies rely on multiple factors viz. number
and type of markers (Chen and Sullivan 2003; Poland and Rife 2012), population
structure (Nakaya and Isobe 2012; Spindel et al. 2015), training population size
(Daetwyler et al. 2008), heritability and architecture of target traits (Zhong et al.
2009; Zhang et al. 2014, 2016) and the relationship between training population and
selection candidates.
Numerous GS models have been proposed to address the diverse requirements
for achieving satisfactory prediction accuracies. Some of the routinely used GS
models include Random Regression Best Linear Unbiased Predictor (RR-BLUP;
Meuwissen et al. 2001; Liu et al. 2008; Zhang et al. 2010), Least Absolute
Shrinkage and Selection Operator (LASSO) (Tibshirani 1996; de los Campos
et al. 2009a), semiparametric strategies (Kinship GAUSS), Bayesian approach
viz. Bayesian Ridge Regression, Bayesian LASSO (de los Campos et al. 2009b;
Legarra et al. 2011), Bayes A (Meuwissen et al. 2001), Bayes B (Meuwissen et al.
2001) and Bayes Cπ (Habier et al. 2011) and machine learning Random Forest
Regression (RFR) (Breiman, 2001), and Support Vector Regression (SVR)
(Drucker et al. 1997). Various comparative accounts have been drawn to assess
the performances of these GS models among different organisms (Moser et al.
2009, Heslot et al. 2012, Resende et al. 2012a, b). Selection of an appropriate GS
model varies from case to case, and hence, multiple models should be considered in
any GS study.
Size of training population is another important factor that has significant impact
on prediction accuracies. Bernardo and Yu (2007) suggested that a minimum size of
the training population to be 100–150 genotypes to obtain the optimum prediction
accuracy. In the case of genetically diverse populations, larger training populations
are required to attain better prediction accuracies (Mujibi et al. 2011). Genetic
relatedness of the individuals in the training and selection populations is known to
affect the accuracies of GS studies (Asoro et al. 2011). Among cattle, GEBVs
estimated within breed were found to be more accurate than the ones estimated
across breeds (Hayes et al. 2009). Price et al. (2010) and Guo et al. (2014)
demonstrated significant reduction in prediction accuracies in structured
populations.
Application of genome-wide markers results in better prediction accuracies
(Meuwissen et al. 2001; Calus and Veerkamp 2007). Higher marker density has
been demonstrated to produce higher genomic prediction accuracy (Zhong et al.
2009; Asoro et al. 2011; Heffner et al. 2011; Poland et al. 2012; Heslot et al. 2013).
Low marker densities in some cases result in lower prediction accuracies, that could
be explained as lower probability of LD between markers and QTLs, because of the
smaller fraction of variation (Solberg et al. 2008). Hickey et al. (2014) reported that
a small number of markers (200–500) and phenotypes (1000) are required in a
closely related biparental population to achieve effective prediction accuracies,
whereas for a population that is unrelated to the selection candidates, a much larger
number of markers and phenotypes are required for the same prediction accuracy. A
large mixed training population set with higher marker density is recommendable to
achieve high prediction accuracies rather than using multiple training populations
representing one germplasm group (Asoro et al. 2011). In another study, De Roos
136 A. Jain et al.
et al. (2009) suggested that a high marker density is required if training and
selection populations are highly divergent.
High-throughput genotyping platforms such as DArT, SNP array and GBS are
being used based on different needs. GBS has been deployed in almost all the crops
in the initial genetic analysis as it provides a low cost option to plant species where
there is no reference genome (Poland et al. 2012). A comparison made by Poland
et al. (2012) using GBS for de novo genotyping of testing populations in case of the
wheat (Triticum aestivum L.) genome showed higher prediction accuracies of
0.3–0.5 in comparison to established marker platforms.
Enhancing the marker numbers while imputing the missing marker data has been
reported to improve in prediction accuracies. For instance, Poland et al. (2012)
showed an improvement of prediction accuracies with the genotyping data set
consisting of 35,000 SNPs with up to 80% missing data points, over the prediction
accuracies estimated from 2000 DArT markers with missing data points up to 2%.
In various studies including maize, wheat, barley and forest trees, a positive
relationship between the trait heritability and prediction accuracies has been
observed (Lorenzana and Bernardo 2009; Albrecht et al. 2011; Heffner et al.
2009, 2011; Grattapaglia et al. 2011; Guo et al. 2012; Combs and Bernardo
2013). In another study, Zhang et al. (2014) established higher prediction accura-
cies for less complex traits. Most of the results discussed here form the basis of
ongoing efforts in legume genomic selection and serve as the guidelines for
strategizing the future efforts. GS efforts in different legumes have been described
below in detail.
6.3 Soybean (Glycine max)
Deployment of GS among legumes first started with improving yield and agro-
nomic traits in soybean. A set of 301 elite breeding lines was genotyped with GBS
and phenotyped for grain yield at multiple locations (Table 6.1) (Jarquı́n et al.
2014). By keeping a randomly selected set of 50 accessions for a validation
population, a positive relationship was observed between the size of training
population and prediction accuracy, which began to plateau at a training population
size of 100; however, it continued to increase until the maximum available size. The
study included the evaluation of three different imputation methods to impute the
missing data for soybean. However, not many differences were obtained using these
imputation methods. Although, random forest imputation produced the highest
accuracies, no significant differences were observed. A high prediction accuracy
(0.64) reflected high potential of GS for yield in soybean (Table 6.1) (Jarquı́n et al.
2014).
Further, exploiting the GAB, genotyping data for 31,045 SNPs on 309 soybean
germplasm accessions were used to estimate the prediction accuracy for seed
weight (SW) (Zhang et al. 2016). Five-fold cross validation (CV) was applied by
randomly assigning 20% of the association panel as validation set and remaining
Table 6.1 Summary of key genomic selection studies in some legume crops
Legume
crop Population Size Marker type Traits GS models Reference
Soybean 301 GBS 1. Grain yield A standard Genomic best linear unbiased predic- Jarquı́n et al.
tion (G-BLUP) model including only additive (2014)
effects, and an extended version of the G-BLUP
model including additive-by-additive effects.
Alfalfa 190 GBS 1. Single harvest Random Regression Best Linear Unbiased Predic- Li et al. (2015)
(10,000 SNPs) biomass tor (RR-BLUP)
2. Total biomass
278 (adapted to two GBS 1. Dry matter yield Support vector regression using linear and Gauss- Annicchiarico
different ian kernel, RR-BLUP, random Forest regression et al. (2015)
environment) and Bayes A, Bayes B and Bayesian lasso,
Pea 372 331 SNP 1. Date of flowering LASSO (least absolute shrinkage ans selection Burstin et al.
2. Number of seeds operator), PLS (partial least squares), SPLS (2015)
per plant (sparse partial least squares), Bayes A, Bayes B
3. Thousand seed and G-BLUP
weight
339 9824 SNPs 1. Date of flowering Kernel partial least squares regression (kPLSR), Tayeh et al.
(GenoPea 13.2 K 2. Number of seeds LASSO, G-BLUP, Bayes A, and Bayes B (2015)
6 Current Status and Prospects of Genomic Selection in Legumes
SNP Array) per plant

3. Thousand seed
weight
Chickpea 320 3000 DArT markers 1. Seed yield RR-BLUP, kinship GAUSS, Bayes Cπ, Bayes B, Roorkiwal et al.
2. 100 seed weight Bayesian LASSO, Random Forest (RF) (2016)
3. Days to 50%
flowering
4. Days to maturity
Groundnut 188 2356 DArT markers 1. Days to flowering RR-BLUP, kinship GAUSS, Bayes Cπ, Bayes B, Pandey et al.
2. Seed weight Baysian LASSO and RF (2014b, 2015)
137
3. Pod yield
138 A. Jain et al.
80% as the training set. Based on the number of SNPs used and the size of training
population, the prediction accuracies were found to vary between 0.75 and 0.87.
Like other studies (Asoro et al. 2011; Jarquin et al. 2014), on size of the training
population, smaller populations resulted in lower prediction accuracies. Another
observation was the prediction accuracy using all 2000 SNPs was found to be same,
even reducing it to 500 SNPs. Higher prediction accuracies were observed com-
pared to Jarquı́n et al. (2014) with same number of markers, similar population size,
and broad sense heritability of traits, pointing towards the impact of genetic
architecture of traits in populations under investigation.
6.4 Alfalfa (Medicago sativa)
Alfalfa is a perennial legume with a long breeding cycle, which limits crop
improvement efforts. Selection cycle duration can be reduced by deploying GS
for complex traits such as yield by using GS for predicting the breeding values
(Li et al. 2015). Prediction accuracies were obtained using phenotyping data for
yield traits during two selection cycles from three locations and using genotyping
data for ~10,000 SNPs (Li et al. 2015). Varying levels of missing values from the
marker data set were used for GS modelling using random forest method for
missing values imputation. Validation of genomic prediction models was
performed by cross validation, in which randomly selected 90% genotypes were
used as training population and 10% was used for testing/validation. Marker data
sets with more missing values resulted in a large number of markers and resulted in
increased prediction accuracies. Prediction accuracies were validated for both the
generation viz. cycle 0 and cycle 1. In individual generation analysis, prediction
accuracies validated within locations were found to be much higher than prediction
accuracies across the locations, possibility due to G E interaction for biomass
yield. Prediction accuracies of 0.43–0.66 for total biomass yield in a synthetic
alfalfa breeding population showed the underlying potential of further application
of GS in other complex traits (Li et al. 2015) (Table 6.1).
In total, 278 elite genotypes adapted to two different environments with a
different genetic base were genotyped using GBS and phenotyped for dry matter
yield of their densely planted half-sib progenies in separate environments
(Annicchiarico et al. 2015). Prediction accuracies were higher using joint SNP
calling in comparison to separate SNP calling for the two data sets. Random forest
was used for missing marker imputation. A comparison of prediction accuracies
within and across populations was performed with the same set of markers, and it
was observed that within-population prediction accuracies were higher than across-
population prediction accuracies, probably due to a high level of intra-population
variation. Results indicated a greater than three-fold higher prediction for yield gain
per unit time though GS in comparison to conventional selection (Annicchiarico
et al. 2015) (Table 6.1).
6.5 Pea (Pisum sativum)
In the case of pea, SNP markers were used to predict the phenotypes using different
statistical methods (Burstin et al. 2015). Phenotyping data for two seasons and
genotyping data generated with 331 SNPs on >350 accessions representing various
cultivars, diverse wild types, landraces, etc. were used to estimate the prediction
accuracies (Table 6.1). To minimize the impact of population structure leading to
spurious associations, authors used the approach recommended by Johnson et al.
(2007). Thousand seed weight (TSW) was predicted better than the beginning of
flowering (BegFlo) and number of seeds per plant (NSeed). During the same year,
they reported deployment of a high-density genotyping platform for GS (Tayeh
et al. 2015). Similarly, genotyping data from the GenoPea 13.2 K SNP Array on a
collection of 339 accessions along with the phenotyping data for TSW, BegFlo and
NSeed were used for estimating genomic prediction values using five different
statistical methods (Tayeh et al. 2015). To estimate the impact of the training
population size over the prediction accuracies, different sizes of training
populations were selected randomly with multiple repetitions; however, the test
set was fixed with 99 accessions. Similarly, to assess the effect of marker density on
prediction accuracies, evenly distributed SNP subsets were selected for estimation.
Of five models considered in the study, four showed equivalent performance,
whereas performance of LASSO was less than others. Another highlight of the
study was that no significant differences were observed whether or not the markers
with low minor allele frequency (MAF) were included. The effect of a reduction in
the size of the training population was reduction in accuracy of the prediction
models (Q2). In addition, reducing the marker density but retaining only a single
marker per unique map position did not affect prediction accuracy. However, a
further reduction in the number of markers led to reduced Q2. Q2 values obtained in
Tayeh et al. (2015) were found to be higher than in Burstin et al. (2015).
6.6 Chickpea (Cicer arietinum)
In case of chickpea, there is only one report coming from ICRISAT about deploying
GS breeding and conducting initial studies of standardizing different GS models
(Roorkiwal et al. 2016). In this context, a training population containing 320 elite
chickpea breeding lines consisting of desi and kabuli seed types, from the Interna-
tional Chickpea Screening Nursery (ICSN), was genotyped using the DArTseq
platform. This platform generated 3000 polymorphic markers. Phenotyping data
were generated for yield and yield-related traits viz. seed yield (SY), 100 seed
weight (SDW), days to 50% flowering (DF) and days to maturity (DM), at two
different locations during two different crop seasons for two different treatments,
that is, rainfed and irrigated conditions. Six different statistical models were used to
calculate prediction accuracies and perform five-fold cross validation to estimate
140 A. Jain et al.
the prediction accuracies by randomly selecting 80% of the lines for the training
population and the remaining 20% as the testing population (Roorkiwal et al. 2016).
A large variation in prediction accuracies were observed among the traits under-
taken in the study, but overall performance of the models were found to be similar
for every trait. The effect of G E interaction was observed in the prediction
accuracies of individual traits. For instance, the best prediction accuracy was
observed for SDW (trait least affected by G x E interaction and treatments, etc.);
however, prediction accuracies were lower for SY trait, which is known to be
affected by G E. The impact of missing marker data and MAF on prediction
accuracies was assessed for 100 seed weight, using nine different combinations of
missing marker data and MAF (including markers in combination with 0%, 10%
and 30% missing data, and 0%, 5% and 10% MAF). The results showed that
the random forest model at 0% missing marker data and 5% MAF combination
had the best prediction accuracy, whereas the Bayes B model with 0% missing
marker data and 10% MAF produced lowest accuracies. This study also assessed
the impact of population structure on GEBV prediction accuracy. Desi and kabuli
seed types were undertaken as separate groups and also grouped together to
calculate prediction accuracies. The results reflected a higher prediction accuracy
using the complete set in comparison to different seed types considered separately,
which might be attributed to a larger population size (Roorkiwal et al. 2016)
(Table 6.1).
6.7 Groundnut (Arachis hypogaea)
In case of groundnut, ICRISAT has taken some initiatives towards deploying GS

breeding and conducting initial studies of standardizing different GS models
(Pandey et al. 2016).While undertaking deployment of GS in groundnut, the
focus of the study was to assess the impact of associated markers on prediction
accuracies for three important traits viz. days to flowering (DF), seed weight
(SW) and pod yield (PY) with different heritabilities (Pandey et al. 2014a, b;
Pandey et al. 2015). Six seasons of phenotyping data for these traits and genotyping
of the reference set with 2356 DArT markers were used for GS analysis (Table 6.1).
When comparing the prediction accuracy for total and associated markers, the
impact of population size and two different approaches were used to estimate the
prediction accuracies. In the first approach, the whole population set was consid-
ered as a training population, and a part of the training population was considered as
validation set to calculate the prediction accuracies. However, in another approach,
the whole population was fractioned into five random smaller sets, of which one set
was used to train the GS model, hence acted as training population, and the rest four
were used as validation sets. Associated markers were compared with using all
markers and the associated marker set showed higher prediction accuracies. How-
ever in a second approach where randomly selected smaller sets were used to
genotype the training population, prediction accuracies obtained with associated
markers were less predictive than all genome-wide markers. Overall, only marginal
differences were observed between the prediction accuracies estimated using total
genome-wide markers by both the approaches. As expected, the traits with higher
heritability showed higher prediction accuracies in comparison to those with lower
heritability. A positive relation between the heritability and prediction accuracies
was observed, supporting similar observations in maize, wheat, barley, etc.
(Lorenzana and Bernardo 2009; Albrecht et al. 2011; Heffner et al. 2011; Guo
et al. 2012; Combs and Bernardo 2013). So far, the lack of a high-throughput
genotyping platform to generate high-density genotyping data has been the major
obstacle in deploying the GS breeding in groundnut. However, the availability of
genome sequences of a diploid progenitor species and 58 K Axiom_Arachis SNP
(Pandey et al. 2017) array during 2016 will further boost the deployment of GS
breeding in groundnut.
6.8 Conclusions
The majority of legume crops lacked the attention of researchers for generating
genomic resources for a longer time compared with cereal crops. Nevertheless, the
speedy development in NGS technologies and assembly methodologies made
generating genomic resources affordable and technically sound over the time.
The legume crops have made much progress from poor resource to highly enriched
genomic resourced crops. This has provided many opportunities to implement
advanced genomic-assisted breeding. GS breeding has demonstrated its great
value to the ongoing conventional breeding programs of cattle and in some plant
species. This approach is gaining attention from other crop breeders including
legumes as it promises greater genetic gain by improving complex traits in less
time with more precision. Seeing the benefits achieved in the maize and wheat
breeding programs, legume crops are now looking forward to deploying GS breed-
ing to address its some of the most complex problems that are the key obstacles in
achieving higher productivity. Selected studies conducted so far in legumes have
suggested the possibility of achieving high prediction accuracies. These prelimi-
nary studies also indicated the potential role of GS in developing superior varieties
with enhanced genetic gain and ability to overcome various stresses, hence ensuring
food security with higher productivity. Currently, the majority of the legume crops
are in the process of deploying GS in their breeding program; however, it will take a
few years for GS to become routine similar to other major crop breeding programs.
References
Albrecht T, Wimmer V, Auinger H, Erbe M, Knaak C et al (2011) Genome-based prediction of

testcross values in maize. Theor Appl Genet 123:339–350
142 A. Jain et al.
Annicchiarico P, Nazzicari N, Li X, Wei Y, Pecetti L et al (2015) Accuracy of genomic selection

for alfalfa biomass yield in different reference populations. BMC Genomics 16:1020
Araújo SS, Beebe S, Crespi M, Delbreil B, González EM et al (2015) Abiotic stress responses in
Legume crops: strategies used to cope with environmental challenges. Crit Rev Plant Sci
34:237–280
Asoro FG, Newell MA, Beavis WD, Scott MP, Jannink J (2011) Accuracy and training population
design for genomic selection on quantitative traits in elite North American oats. The Plant
Genome 4:132–144
Ates D, Sever T, Aldemir S, Yagmur B, Temel HY et al (2016) Identification QTLs Controlling
Genes for Se Uptake in Lentil Seeds. PLOS ONE 11(4): e0154054
Bernardo R, Charcosset A (2006) Usefulness of gene information in marker-assisted recurrent
selection: a simulation appraisal. Crop Sci 46:614–621
Bernardo R, Yu J (2007) Prospects for genome-wide selection for quantitative traits in maize. Crop
Sci 47:1082–1090
Bertioli DJ, Cannon SB, Froenicke L, Huang G, Farmer AD et al (2016) The genome sequences of
Arachis duranensis and Arachis ipaensis, the diploid ancestors of cultivated peanut. Nat Genet
48:438–446
Bohra A, Pandey MK, Jha UC, Singh B, Singh IP et al (2014) Genomics-assisted breeding in four
major pulse crops of developing countries: present status and prospects. Theor Appl Genet
127:1263–1291
Bordat A, Savois V, Nicolas M, Salse J, Chauveau A et al (2011) Translational genomics in
legumes allowed placing in silico 5460 unigenes on the pea functional map and identified
candidate genes in Pisum sativum L. Genes Genome Genet 1:93–103
Breiman L (2001) Random forests. Mach Learn 45:5–32. doi:10.1023/A:1010933404324
Burstin J, Salloignon P, Chabert-Martinello M, Magnin-Robert JB, Siol M et al (2015) Genetic
diversity and trait genomic prediction in a pea diversity panel. BMC Genomics 16:105
Calus MPL, Veerkamp RF (2007) Accuracy of breeding values when using and ignoring the
polygenic effect in genomic breeding value estimation with a marker density of one SNP per
cM. J Ani Breed Genet 124:362–368
Chen X, Sullivan PF (2003) Single nucleotide polymorphism genotyping: biochemistry, protocol,
cost and throughput. Pharmacogenomics J 3:77–96
Chen X, Li H, Pandey MK, Yang Q, Wang X et al (2016) Draft genome of the peanut A-genome
progenitor (Arachis duranensis) provides insights into geocarpy, oil biosynthesis, and aller-
gens. Proc Nat Acad Sci 113:6785–6790
Combs E, Bernardo R (2013) Accuracy of genomewide selection for different traits with constant
population size, heritability and number of markers. The Plant Genome 6:1
Cottage A, Gostkiewicz K, Thomas JE, Borrows R, Torres AM et al (2012) Heterozygosity and
diversity analysis using mapped SNPs in a faba bean inbreeding programme. Mol Breed
30:1799–1809
disease using a genome-wide approach. PLoS One 3:e3395
Davey JW, Hohenlohe PA, Etter PD, Boone JQ, Catchen JM et al (2011) Genome-wide genetic
marker discovery and genotyping using next-generation sequencing. Nat Rev Genet
12:499–510
de los Campos G, Gianola D, GJM R (2009a) Reproducing kernel Hilbert spaces regression: a
general framework for genetic evaluation. J Anim Sci 87:1883–1887
de los Campos G, Naya H, Gianola D, Crossa J, Legarra A et al (2009b) Predicting quantitative
traits with regression models for dense molecular markers and pedigree. Genetics 182:375–385
de Roos APW, Hayes BJ, Goddard ME (2009) Reliability of genomic breeding values across
multiple populations. Genetics 183:1545–1553. doi:10.1534/genetics.109.104935
Deokar AA, Ramsay L, Sharpe AG, Diapari M, Sindhu A et al (2014) Genome wide SNP
identification in chickpea for use in development of a high density genetic map and improve-
ment of chickpea reference genome assembly. BMC Genomics 15:708
Deulvot C, Charrel H, Marty A, Jacquin F, Donnadieu C et al (2010) Highly-multiplexed SNP

genotyping for genetic mapping and germplasm diversity studies in pea. BMC Genomics
11:468
Drucker H, Burges CJC, Kaufman L, Smola AJ, Vapnik V (1997) Support vector regression
machines. Adv Neural Info Process Syst 9:155–161
Duarte J, Rivière N, Baranger A et al (2014) Transcriptome sequencing for high throughput SNP
development and genetic mapping in pea. BMC Genomics 15:126
Egbadzor KF, Ofori K, Yeboah M, Aboagye LM, Opoku-Agyeman MO et al (2014) Diversity in
113 cowpea [Vigna unguiculata (L) Walp] accessions assessed with 458 SNP markers.
Springer Plus 3:541
Fedoruk M (2013) Linkage and association mapping of seed size and shape in lentil. Thesis
(Masters of Science), University of Saskatchewan, Saskatoon
Gautami B, Pandey MK, Vadez V, Nigam SN, Ratnakumar P et al (2012) Quantitative trait locus
analysis and construction of consensus genetic map for drought tolerance traits based on three
recombinant inbred line populations in cultivated groundnut (Arachis hypogaea L.) Mol Breed
30:757–772
Graham PH, Vance CP (2003) Legume crops: importance and constraints to greater use. Plant
Physiol 131:872–877
Grattapaglia D, Resende MDV, Resende MR, Sansaloni CP, Petroli CD et al (2011) Genomic
selection for growth traits in eucalyptus: accuracy within and across breeding populations.
BMC Proc 5:O16. doi:10.1186/1753-6561-5-S7-O16
Gujaria N, Kumar A, Dauthal P, Dubey A, Hiremath P et al (2011) Development and use of genic
molecular markers (GMMs) for construction of a transcript map of chickpea (Cicer arietinum
L.) Theor Appl Genet 122:1577–1589
Guo Z, Tucker D, Lu J, Kishore V, Gay G (2012) Evaluation of genome-wide selection efficiency
in maize nested association mapping populations. Theor Appl Genet 124:261–275
Guo Z, Tucker DM, Basten CJ, Gandhi H, Ersoz E et al (2014) The impact of population structure
on genomic prediction in stratified populations. Theor Appl Genet 127:749–762
Habier D, Fernando RL, Kizilkaya K, Garrick DJ (2011) Extension of the Bayesian alphabet for
genomic selection. BMC Bioinformatics 12:186
Hascoët E, Jaminon O, Devaux C, Blassiau C, Bahrman N, Bochard A-M et al (2014) Towards fine
mapping of frost tolerance QTLs in pea, in 2nd PeaMUST Annual Meeting (Dijon)
Hayes B, Bowman P, Chamberlain A, Goddard M (2009) Invited review: genomic selection in
dairy cattle: progress and challenges. J Dairy Sci 92:433–443
Heffner EL, Me S, Jannink JL (2009) Genomic selection for crop improvement. Crop Sci 49:1–12
Heffner EI, Jannink JL, Iwata H, Souza E, Sorrells ME (2011) Genomic selection accuracy for
grain quality traits in biparental wheat populations. Crop Sci 51:2597–2606
Heslot N, Yang HP, Sorrells ME, Jannink JL (2012) Genomic selection in plant breeding: a
comparison of models. Crop Sci 52:146–160
Heslot N, Rutkoski J, Poland J, Jannink JL, Sorrells ME (2013) Impact of marker ascertainment
bias on genomic selection accuracy and estimates of genetic diversity. PLoS One 8:e74612
Hickey JM, Dreisigacker S, Crossa J, Hearne S, Babu R et al (2014) Evaluation of genomic
selection training population designs and genotyping strategies in plant breeding programs
using simulation. Crop Sci 54:1476–1488
Hiremath PJ, Farmer A, Cannon SB, Woodward J, Kudapa H et al (2011) Large-scale
transcriptome analysis in chickpea (Cicer arietinum L.), an orphan legume crop of the semi-
arid tropics of Asia and Africa. Plant Biotechnol J 9:922–931. doi:10.1111/j.1467-7652.2011.
00625.x
Hiremath PJ, Kumar A, Penmetsa RV, Farmer A, Schlueter JA et al (2012) Large-scale develop-
ment of cost-effective SNP marker assays for diversity assessment and genetic mapping in
chickpea and comparative mapping in legumes. Plant Biotechnol J 10:716–732
Hospital F (2005) Selection in backcross programmes. Philos Trans Roy Soc Lond B Biol Sci
360:1503–1511
144 A. Jain et al.
Huynh BL, Close TJ, Roberts PA, Hu Z, Wanamaker S et al (2013) Gene pools and the genetic
architecture of domesticated cowpea. The Plant Genome 6:3
Jaganathan D, Thudi M, Kale S, Azam S, Roorkiwal M et al (2015) Genotyping-by-sequencing
based intra-specific genetic map refines a “QTL-hotspot” region for drought tolerance in
chickpea. Mol Gen Genomics 290:559–571
Jain M, Misra G, Patel RK, Priya P, Jhanwar S et al (2013) A draft genome sequence of the pulse
crop chickpea (Cicer arietinum L.) Plant J 74:715–729. doi:10.1111/tpj.12173
Janila P, Pandey MK, Shasidhar Y, Variath MT, Sriswathi M et al (2016) Molecular breeding for
introgression of fatty acid desaturase mutant alleles (ahFAD2A and ahFAD2B) enhances oil
quality in high and low oil containing peanut genotypes. Plant Sci 242:203–213
Jarquı́n D, Kocak K, Posadas L, Hyma K, Jedlicka J et al (2014) Genotyping by sequencing for
genomic prediction in a soybean breeding population. BMC Genomics 15:740
Johnson WE, Li C, Rabinovic A (2007) Adjusting batch effects in microarray expression data
using empirical bayes methods. Biostatistics 8:118–127. doi:10.1093/biostatistics/kxj037
Kang YJ, Kim SK, Kim MY, Lestari P, Kim KH et al (2014) Genome sequence of mungbean and
insights into evolution within Vigna species. Nat Commun 5:5443
Kaur S, Cogan NO, Stephens A, Noy D, Butsch M et al (2014a) EST-SNP discovery and dense
genetic mapping in lentil (Lens culinaris Medik.) enable candidate gene selection for boron
tolerance. Theor Appl Genet 127:703–713
Kaur S, Kimber RBE, Cogan NOI, Materne M, Forster JW et al (2014b) SNP discovery and high-
density genetic mapping in faba bean (Vicia faba L.) permits identification of QTLs for
ascochyta blight resistance. Plant Sci 217–218:47–55
Khera P, Upadhyaya HD, Pandey MK, Roorkiwal M, Sriswathi M et al (2013) SNP-based genetic
diversity in the reference set of peanut (Arachis spp.) by developing and applying cost-
effective KASPar genotyping assays. Plant Genome 6:1–11
Kim KH, Kim MY, Van K, Moon JK, Kim DH et al (2008) Marker-assisted foreground and
background selection of near isogenic lines for bacterial leaf pustule resistant gene in soybean.
J Crop Sci Biotechnol 11:263–268
Kumar S, Banks TW, Cloutier S (2012) SNP discovery through next-generation sequencing and its
applications. Int J Plant Genomics 15
Kumar V, Rani A, Rawal R, Mourya V (2015) Marker assisted accelerated introgression of null
allele of kunitz trypsin inhibitor in soybean. Breed Sci 65:447–452
Lavaud C, Lesne A, Piriou C, Le Roy G, Boutet G et al (2015) Validation of QTL for resistance to
Aphanomyces euteiches in different pea genetic backgrounds using near-isogenic lines. Theor
Appl Genet 128:2273–2288
Lee YG, Jeong N, Kim JH, Lee K, Kim KH et al (2015) Development, validation and genetic
analysis of a large soybean SNP genotyping array. Plant J 81:625–636
Legarra A, Robert-Granie P, Croiseau G, Guillaume F, Fritz S (2011) Improved LASSO for
genomic selection. Genet Res 93:77–87
Li X, Wei Y, Acharya A, Hansen JL, Crawford JL et al (2015) Genomic prediction of biomass
yield in two selection cycles of a tetraploid alfalfa breeding population. Plant Genome 8
Liu XQ, Rong JY, Liu XY (2008) Best linear unbiased prediction for linear combinations in
general mixed linear models. J Multivariate Analysis 99:1503–1517
selection in biparental plant populations. Theor Appl Genet 120:151–161
Lucas MR, Ehlers JD, Huynh BL, Diop NN, Roberts PA et al (2013) Markers for breeding heat-
tolerant cowpea. Mol Breed 31:529–536
Moser G, Tier B, Crump RE, Khatkar MS, Raadsma HW (2009) A comparison of five methods to
predict genomic breeding values of dairy bulls from genome-wide SNP markers. Genet Sel
Evol 41:56
Mujibi FD, Nkrumah JD, Durunna ON, Stothard P, Mah J et al (2011) Accuracy of genomic
breeding values for residual feed intake in crossbred beef cattle. J Anim Sci 89:3353–3361
Mu~noz-Amatriaı́n M, Mirebrahim H, Xu P, Wanamaker SI, Luo M et al (2016) Genome resources
for climate-resilient cowpea, an essential crop for food security. Plant J. doi:https://doi.org/10.
1101/059261
Nakaya A, Isobe SN (2012) Will genomic selection be a practical method for plant breeding? Ann
Bot 110:1303–1316
Pandey MK, Rathore A, Das RR, Khera P, Upadhyaya HD et al (2014a) Selection of appropriate
genomic selection model in an unstructured germplasm set of peanut (Arachis hypogaea L.).
6th international Food Legumes Research conference & 7th international conference on
Legume Genetics and Genomics on 7–11 July 2014, Saskatoon
Pandey MK, Upadhyaya HD, Rathore A, Vadez V, Sheshshayee MS et al (2014b) Genome-wide
association studies for 50 agronomic traits in peanut using the ‘reference set’ comprising
300 genotypes from 48 countries of semi-arid tropics of the world. PLoS One 9:e113326
Pandey MK, Agarwal G, Rathore A, Janila P, Upadhyaya HD, et al. (2015). Development of high
density 60K “Axiom_Arachis” SNP Chip and optimization of genomic selection model for
enhancing breeding efficiency in peanut. Proceedings of 8th international conference of the
Peanut Research Community on “Advances in Arachis through Genomics and Biotechnology”,
Brisbane, 5–9 Nov 2015
Pandey MK, Roorkiwal M, Singh VK, Ramalingam A, Kudapa H et al (2016) Emerging genomic
tools for legume breeding: current status and future prospects. Front Plant Sci 7
Pandey MK, Agarwal G, Kale SM, Clevenger J, Nayak SN et al (2017) Development and
evaluation of a high density genotyping ‘Axiom_Arachis’ array with 58K SNPs for accelerat-
ing genetics and breeding in groundnut. Nat Sci Rep 7:40577. doi:10.1038/srep40577
Poland J, Rife TW (2012) Genotyping-by-sequencing for plant breeding and genetics. Plant
Genome 5:92–102
Poland J, Endelman J, Dawson J, Rutkoski J, Wu S et al (2012) Genomic selection in wheat
breeding using genotyping-by-sequencing. Plant Genome 5:103–113
Power JF (1987) Legume crops: their potential role in agricultural production. Am J Alt Agri
2:69–73
Price AL, Zaitlen NA, Reich D, Patterson N (2010) New approaches to population stratification in
genome-wide association studies. Nat Rev Genet 11:459–463
Rebello CJ, Greenway FL, Finley JW (2014) A review of the nutritional value of legumes and their
effects on obesity and its related co-morbidities. Obesity Rev 15:392–407
Resende MDV, Resende MFR, Sansaloni CP, Petroli CD, Missiaggia AA et al (2012a) Genomic
selection for growth and wood quality in Eucalyptus: capturing the missing heritability and
accelerating breeding for complex traits in forest trees. New Phytol 194:116–128
Resende MFR, Munoz P, Resende MDV, Garrick DJ, Fernando RL et al (2012b) Accuracy of
genomic selection methods in a standard data set of loblolly pine (Pinus taeda l.) Genetics
190:1503–1510
Roorkiwal M, Rathore A, Das RR, Singh MK, Jain A, Srinivasan S, Gaur PM, Chellapilla B,
Tripathi S, Li Y, Hickey JM, Lorenz A, Sutton T, Crossa J, Jannink J-L, Varshney RK (2016)
Genome-Enabled prediction models for yield related traits in chickpea. Front Plant Sci 7
Roorkiwal M, Sawargaonkar SL, Chitikineni A, Thudi M, Saxena RK et al (2013) Single
nucleotide polymorphism genotyping for breeding and genetics applications in chickpea and
pigeonpea using the BeadXpress platform. Plant Genome 6
Roorkiwal M, Von Wettberg EJ, Upadhyaya HD, Warschefsky E, Rathore A et al (2014)
Exploring germplasm diversity to understand the domestication process in Cicer spp. using
SNP and DArT markers. PLoS One 9(7):e102016
Rubiales D, Mikic A (2015) Introduction: legumes in sustainable agriculture. Crit Rev Plant Sci
34:2–3
Sato S, Nakamura Y, Kaneko T, Asamizu E, Kato T et al (2008) Genome structure of the legume,
Lotus japonicus. DNA Res 15:227–239
146 A. Jain et al.
Saxena RK, Penmetsa RV, Upadhyaya HD, Kumar A, Carrasquilla-Garcia N et al (2012) Large-
scale development of cost-effective single-nucleotide polymorphism marker assays for genetic
mapping in pigeonpea and comparative mapping in legumes. DNA Res 19:449–461
Schmutz J, Cannon SB, Schlueter J, Ma J, Mitros T et al (2010) Genome sequence of the
palaeopolyploid soybean. Nature 463:178–183
Sharpe AG, Ramsay L, Sanderson LA, Fedoruk MJ, Clarke WE et al (2013) Ancient orphan crop
joins modern era: gene-based SNP discovery and mapping in lentil. BMC Genomics 14:192
Solberg TR, Sonesson AK, Woolliams JA (2008) Genomic selection using different marker types
and densities. J Anim Sci 86(10):2447–2454
Spindel J, Begum H, Akdemir D, Virk P, Collard B et al (2015) Genomic selection and association
mapping in rice (Oryza sativa): effect of trait genetic architecture, training population com-
position, marker number and statistical model on accuracy of rice genomic selection in elite,
tropical rice breeding lines. PLoS Genet 11:e1005350
Tarawali G, Ogunbile OA (1995) Legumes for sustainable food production in semi-arid savan-
nahs. ILEIA Newslett 11(4):18–23
Tayeh N, Aluome C, Falque M, Jacquin F, Klein A et al (2015) Development of two major
resources for pea genomics: the GenoPea 13.2 K SNP Array and a high-density, high-
resolution consensus genetic map. Plant J 84:1257–1273
Tibshirani R (1996) Regression shrinkage and selection via the lasso. J Roy Stat Soc Series B
58:267–288
Varshney RK, Chen W, Li Y, Bharti AK, Saxena RK et al (2012) Draft genome sequence of
pigeonpea (Cajanus cajan), an orphan legume crop of resource-poor farmers. Nat Biotechnol
30:83–89
Varshney RK, Mohan SM, Gaur PM, Gangarao NVPR, Pandey MK et al (2013a) Achievements
and prospects of genomics-assisted breeding in three legume crops of the semi-arid tropics.
Biotechnol Adv 31:1120–1134
Varshney RK, Song C, Saxena RK, Azam S, Yu S et al (2013b) Draft genome sequence of
chickpea (Cicer arietinum) provides a resource for trait improvement. Nat Biotechnol
31:240–246. doi:10.1038/nbt.2491
Varshney RK, Gaur PM, Chamarthi SK, Krishnamurthy L, Tripathi S et al (2013c) Fast-track
introgression of “QTL-Hotspot” for root traits and other drought tolerance traits in JG 11, an
elite and leading variety of chickpea. Plant Genome 6:3. doi:10.3835/plantgenome2013.07.
0022
Varshney RK, Mohan SM, Gaur PM, Chamarthi SK, Singh VK et al (2014) Marker-assisted
backcrossing to introgress resistance to Fusarium wilt (FW) race 1 and Ascochyta blight
(AB) in C 214, an elite cultivar of chickpea. Plant Genome. doi:10.3835/plantgenome2013.
10.0035
Verma S, Gupta S, Bandhiwal N, Kumar T, Bharadwaj C et al (2015) High-density linkage map
construction and mapping of seed trait QTLs in chickpea (Cicer arietinum L.) using
genotyping-by-sequencing (GBS). Sci Rep 5:17512
Wang J, Chu S, Zhang H, Zhu Y, Cheng H et al (2016) Development and application of a novel
genome-wide SNP array reveals domestication history in soybean. Sci Rep 6
Xu P, Wu XH, Wang BG, Luo J, Liu YH et al (2012) Genome wide linkage disequilibrium in
Chinese asparagus bean (Vigna unguiculata ssp. sesquipedialis) germplasm: implications for
domestication history and genome wide association studies. Heredity 109:34–40
Yang H, Tao Y, Zheng Z, Zhang Q, Zhou G et al (2013) Draft genome sequence, and a sequence-
defined genetic linkage map of the legume crop species Lupinus angustifolius L. PLoS One 8:
e64799
Young ND, Debellé F, Oldroyd GE, Geurts R, Cannon SB et al (2011) The Medicago genome
provides insight into the evolution of rhizobial symbioses. Nature 480:520–524
Zhang Z, Liu J, Ding X, Bijma P, de Koning D-J et al (2010) Best linear unbiased prediction of
genomic breeding values using a trait-specific marker-derived relationship matrix. PLoS One
5:e12648. doi:10.1371/journal.pone.0012648
Zhang Z, Ober U, Erbe M, Zhang H, Gao N et al (2014) Improving the accuracy of whole genome
prediction for complex traits using the results of genome wide association studies. PLoS One 9:
e93017
Zhang J, Song Q, Cregan PB, Jiang GL (2016) Genome-wide association study, genomic predic-
tion and marker-assisted selection for seed weight in soybean (Glycine max). Theor Appl Genet
129:117–130
Zhong S, Dekkers JC, Fernando RL, Jannink JL (2009) Factors affecting accuracy from genomic
selection in populations derived from multiple inbred lines: a barley case study. Genetics
182:355–364
Zhu S, Walker DR, Warrington CV, Parrott WA, All JN et al (2007) Registration of four soybean
germplasm lines containing defoliating insect resistance QTLs from PI 229358 introgressed
into ‘Benning’. J Plant Reg 1:162–163
Chapter 7
Genomic Selection in Hybrid Breeding
Albert Wilhelm Schulthess, Yusheng Zhao, and Jochen C. Reif
Abbreviations
BLUP Best linear unbiased prediction

e-Bayes Empirical Bayes method
GCA General combining ability
GS Genomic selection
LD Linkage disequilibrium
MAS Marker assisted selection
PS Phenotypic selection
RE Relative efficiency
REML Restricted maximum likelihood
RKHS Reproducing kernel Hilbert space
RR-BLUP Ridge regression best linear unbiased prediction
RRS Recurrent reciprocal selection
SCA Specific combining ability
SNP Single nucleotide polymorphism
W-BLUP Weighted best linear unbiased prediction
7.1 Introduction
A “hybrid variety” will be understood as the offspring of a controlled cross of two

or more different (inbred or not) genotypes (Becker 2011). The ultimate goal of
hybrid breeding is the exploitation of the phenomenon known as “heterosis”
A.W. Schulthess • Y. Zhao • J.C. Reif (*)

Department of Breeding Research, Leibniz Institute of Plant Genetics and Crop Plant Research
(IPK), D-06466, Gatersleben, Germany
e-mail: schulthess@ipk-gatersleben.de; zhao@ipk-gatersleben.de; reif@ipk-gatersleben.de

DOI 10.1007/978-3-319-63170-7_7
150 A.W. Schulthess et al.
(Whitford et al. 2013), in which the performance of hybrids is superior to the mean
of its parents (Bernardo 2010; Falconer and Mackay 1996). However, there are
additional reasons why hybrid breeding is preferred to line breeding (Longin et al.
2012):
(i) Hybrids have greater yield stability, which is a major advantage for agriculture
in marginal environments.
(ii) Heterozygosity allows the potential combination of dominant major genes in
the hybrid genotype.
(iii) Hybrids offer a built-in plant variety protection system by means of inbreeding
depression when growing farmed-saved seeds.
In the last years, the plant breeding community started to look at genomic
selection (GS) as a promising tool to reduce the costs and to accelerate plant
breeding programs (Desta and Ortiz 2014; Jannink et al. 2010; Zhao et al.
2014b). The main objective of this chapter is to explain the basic concepts of
hybrid breeding and to integrate them with methods of GS. Thereby, it is expected
that these concepts and methods allow the reader to understand the philosophy
underlying GS in hybrid breeding.
7.2 Basic Concepts Relevant to Hybrid Breeding
Even though heterozygosity does not necessarily imply the occurrence of domi-
nance, heterozygosity is fully required for its existence (Bernardo 2010). This
contrasts with the situation observed for fully inbred genotypes used in line
breeding, in which heterozygosity is practically residual (Bos and Caligari 2008)
and, as a result, dominance effects are expected to be negligible or absent within
this particular system. Therefore, dominance will be considered as a particular
feature of hybrid breeding and will receive special attention in the following
sections. Furthermore, heterosis is of special interest to hybrid breeding, and thus,
basic concepts related to this topic will be presented. Finally, the breeding concepts
of combining ability and heterotic groups will be introduced to understand the
philosophy of hybrid breeding.
7.2.1 Dominance
We will first consider a single-locus model to explain the concept of dominance

(Fig. 7.1). Given two homozygous but contrasting genotypes for locus A, coded as
A1A1 and A2A2, the genotypic values, i.e., the effect of the genotype on the
phenotype, of them and their cross A1A2 can be denoted as: MP a, MP þ a, and
MP þ d, respectively, where MP corresponds to the average of the phenotypes of
parents with genotypes A1A1 and A2A2. The value a stands for the additive effect at
7 Genomic Selection in Hybrid Breeding 151
Fig. 7.1 Schematic representation of the dominance effect (d ) at locus A according to different
levels of dominance: no dominance (d ¼ 0), incomplete dominance (0 < d < a), complete domi-
nance (d ¼ a) and overdominance (d > a). A1A1 and A2A2 represent two homozygous but
contrasting parents at locus A, whereas A1A2 corresponds to the offspring of the cross between
them. MP stands for the mid-parent value between parents A1A1 and A2A2, and a is the difference
between parent A1A1 or A2A2 and the MP value at the phenotypic scale
locus A and is defined as half the difference between the genotypic values of A1A1
and A2A2. When dominance is present at locus A, the value d will differ from zero,
being 0 < d < a, d ¼ a or d > a, in the cases of incomplete dominance, complete
dominance and overdominance, correspondingly (Falconer and Mackay 1996).
Nevertheless, when crosses are performed, the unit of inheritance corresponds to
an allele and not to the genotype itself; therefore, an allelic value should be defined.
The average effect or value of an allele is defined as the mean of individuals
(expressed as a deviation from the population mean) that inherited that particular
allele provided that the second allele was inherited at random (Falconer and
Mackay 1996). In consequence, the genotypic value (g12) of the cross
A1A1A2A2 is denoted as
g12 ¼ μ þ α1 þ α2 þ δ12 , ð7:1Þ
where μ refers to the mean of population in Hardy-Weinberg equilibrium; α1 and α2

are the allele effects of the genes inherited from parents A1A1 and A2A2, respec-
tively; and δ12 represents a residual value, which cannot be explained by the
average allelic effects. This residual term δ12 will be denoted as the dominance
deviation (Falconer and Mackay 1996).
7.2.2 Heterosis
Quantitative geneticists use the term heterosis to make reference to the superiority
of a hybrid over the mean of its parents, a term known as mid-parent heterosis. In
addition, plant breeders tend to use the term better-parent heterosis, in which the
economic advantage of a hybrid is defined as the superior performance over both of
its parents (Becker 2011). Furthermore, in crops such as wheat or barley, where line
breeding historically played a major role compared with hybrid breeding, plant
breeders often use the concept of commercial heterosis, in which the superiority of
the hybrids is judged based on comparison(s) with the best available inbred line(s)
(Longin et al. 2012). Nevertheless, in this chapter, the term “heterosis” will be
understood, if not explicitly stated otherwise, as midparent heterosis.
The genetic causes of heterosis are still a topic of debate in quantitative genetics.
However, there are two main hypotheses that have been proposed to explain this
phenomenon (Bernardo 2010; Whitford et al. 2013):
(i) Dominance hypothesis:
Hybrids are expected to be superior to their homozygous parents because of
the masking of unfavorable recessive alleles in the heterozygous genotype.
Then, at a single locus level, the heterozygote would be expected to not exceed
the genotypic value of the better parent (0 < d a in Fig. 7.1). Nonetheless, it
is very likely that neither of the homozygous parents carry all positive alleles at
all loci for a particular polygenic trait, implying that the hybrid could be
superior to both parents (Bruce 1910; Collins 1921; Jones 1917; Keeble and
Pellew 1910).
(ii) Over-dominance hypothesis:
Hybrids are expected to be superior to their parents because of the inherent
superiority of the heterozygous genotype over both homozygotes (d > a in
Fig. 7.1). In consequence, a single locus would be enough to explain heterosis
(Crow 1948; East 1936; Hull 1945).
Last but not least, even in the absence or with very low degrees of dominance
effects, a situation which could be expected for autogamous species, heterosis could
be present for a particular trait (Bernardo 2010; Whitford et al. 2013). In this case,
epistasis would play an important role in heterosis (Richey 1942; Schnell and
Cockerham 1992). It is likely that a mixture of the above-mentioned mechanisms
ultimately underlies heterosis (Whitford et al. 2013).
7.2.3 Combining Abilities
Once it is known that heterosis guarantees the successful development of commer-

cial hybrid genotypes for a particular species, plant breeders are in general no more
interested in heterosis itself, but rather in the performance of hybrid genotypes.
In this sense, hybrid breeders’ efforts and resources will be completely allocated to
the development of hybrids with superior performance, independently of whether
the superior performance is due to heterosis between the parents or because the
parents have a high per se performance (Bernardo 2010).
Hybrid crop studies have shown for complex traits such as grain yield that hybrid
performance cannot be predicted with high accuracy using the per se performance
of the parents (for review, see Becker 2011 and Hallauer et al. 2010). Therefore,
plant breeders normally evaluate and select good parents based on their perfor-
mance as parents of hybrids, which is often referred to as the combining ability of
the parent lines (Hallauer et al. 2010). Given a pool of parent lines to select, the
hybrid performances of their crosses are evaluated. Then, the mean hybrid perfor-
mance of a particular cross Fi Mj (μFi Mj ) can be expressed using the combining
abilities of parent lines as follows (Bernardo 2010; Falconer and Mackay 1996):
μFi Mj ¼ μ þ GCAFi þ GCAMj þ SCAFi Mj , ð7:2Þ
where μ is the mean of all hybrids, whereas GCA and SCA correspond to the
general and specific combining abilities of parents, respectively. The GCAFi and
GCAMj are the mean values expressed as deviations from μ of all F1 hybrids having
Fi or Mj as one of the parents, correspondingly. However, even ignoring nongenetic
sources of error, there would be a remaining proportion of variability for the hybrid
performance, which could not be explained by the GCA of both parents. This last
term, referred to as SCAFi Mj , measures the interaction between parents Fi and Mj
which cannot be accounted by the main effects of their GCAs. The combining
abilities are assumed as independent from each other. Hence, assuming no error
variance, the total variance between hybrids can be decomposed as
σ 2GCAF þ σ 2GCAM þ σ 2SCA , if parent lines conform groups by sex or factors (denoted
as F for female or M for male) or alternatively as 2σ 2GCA þ σ 2SCA , if they do not
configure any kind of groups (Falconer and Mackay 1996).
GCA is often interpreted as the influence of additive effects, whereas SCA as an
indication of genes having dominance and epistatic effects (Hallauer et al. 2010).
Moreover, provided that both parents are completely homozygous inbred lines and
assuming the absence of epistatic effects, σ 2GCA and σ 2SCA are equal to the variances
due to α (σ 2A ) and δ effects (σ 2D ), respectively (Wricke and Weber 1986).
7.2.4 Heterotic Groups and Patterns
A heterotic group is defined as a group of genotypes that display similar combining

ability and heterotic response when crossed with genotypes from other genetically
distinct germplasm groups. In addition, a pair of specific heterotic groups with a
high hybrid performance in their cross will conform a heterotic pattern (Melchinger
and Gumber 1998). It has been suggested that using genetically divergent
populations is relevant for the establishment of heterotic groups and patterns.

Genetically diverse heterotic groups do not only allow the maximum exploitation
of heterosis and hybrid performance but also lead to a lower σ 2D =σ 2A ratio or,
equivalently, to a lower σ 2SCA =σ 2GCA proportion. The latter implies that hybrid
performance could be predicted using Eq. (7.2) only relying on GCAs of parents
(Reif et al. 2007). Furthermore, because combining ability can be exploited in a
recurrent fashion, breeding efforts within a hybrid breeding program can then be
allocated to the selection of the best parents for each heterotic group within a specific
and previously identified heterotic pattern mainly based on their GCA (Hallauer et al.
2010). This breeding method is known as recurrent reciprocal selection (RRS).
Briefly, provided that at least two heterotic groups or pools are available, genotypes
to be tested of one pool are testcrossed with a small number of random sampled
genotypes that belong to the opposite pool. Therefore, a small number of crosses for
each tested genotype are generated. Later, all seeds pertaining to crosses of a
particular tested parent are harvested and bulked together as a single progeny.
Thus, each parent will be represented by its corresponding progeny in field trials
during the next season. Then, the best parent lines of each pool are recognized and
selected based on the performance of each progeny. Subsequently, selected parents
are inter-mated within each pool and serve as base material for their respective
heterotic group during the next breeding cycle. This whole process is repeated in
parallel for each considered pool (Comstock et al. 1949). The RRS method has been
widely and successfully used for hybrid breeding in crops such as maize, where
heterotic groups and patterns were empirically developed by observing which crosses
produced superior hybrid performance and which do not (Tracy and Chandler 2006).
7.3 GS for Hybrid Genotypes
7.3.1 Cross-Validated Prediction Accuracy of GS
Studies on GS often mix the concepts of prediction ability (or predictability) and
prediction accuracy. Nevertheless, prediction ability is expressed as the correlation
between genomic predictions and observed phenotypes, whereas prediction accu-
racy is generally defined as the prediction ability divided by h (the square root of the
heritability h2) for the trait being predicted (Lorenzana and Bernardo 2009;
Riedelsheimer et al. 2012; Zhao et al. 2012b). In this sense, the prediction accuracy
value is interpreted as prediction ability for a trait with heritability equal to 1, and,
in consequence, prediction accuracies are expected to be higher than prediction
ability values. Moreover, because prediction accuracies provide an estimate of the
genotypic correlation, they are more relevant for the estimation of the effects of
indirect selection by means of GS (Zhao et al. 2013b). However and for simplicity,
both terms will be indistinctly used in the present chapter, unless the contrary is
stated.
Predictabilities of GS are overestimated when the same genotypes used for the
estimation of marker effects are considered for prediction (Krchov et al. 2015; Xu
et al. 2014), even if the observed phenotypes used for the computation of pre-
dictabilities are different from those considered for marker effects estimation
(Krchov et al. 2015). Such approaches to compute predictability do not properly
mimic the situation faced in practice by plant breeders: genomic models will be
trained with an estimation set of genotypes for which phenotypic and genomic data
are available and predictions will be obtained for a group of genotyped, but not
phenotyped, selection candidates, which are in principle independent from the
estimation set. Therefore, validation of the predictability is crucial to show the
actual potential of GS in plant breeding and can be efficiently achieved by means of
cross-validations (Hjorth 1994). In k-fold cross-validation, for instance, the popu-
lation with available genomic and phenotypic data is divided in k subgroups of
similar size. Then, the first k 1 subgroups are used to predict the effects of
markers, and the genotypes included in the kth subgroup are predicted and compared
with their observed values. This process can be iteratively repeated to obtain robust
estimates for the cross-validated prediction accuracy of GS.
7.3.1.1 Relatedness Plays a Major Role in Determining the Cross-

Validated Prediction Accuracy of GS in Hybrid Breeding
From simulation and experimental plant data studies, it is well known that related-
ness between estimation and validation sets influences the prediction ability of GS
(Gowda et al. 2014; Habier et al. 2007; Meuwissen 2009; Meuwissen et al. 2001;
Mirdita et al. 2015; Technow et al. 2012, 2014; Zhao et al. 2013b, 2015). In this
sense, the more related these two sets are, the higher would be the predictability. In
addition, relatedness will be present at different levels of a plant breeding program,
and, in consequence, this should be taken into account at the moment of performing
GS and interpreting predictability levels. The current section aims to illustrate this
point in detail.
Relatedness and Its Implications on the Predictabilities of GS Within Hybrid

Breeding Programs
Different levels of relatedness can be found within the plant material used in a
hybrid breeding program. This is schematically represented in Fig. 7.2 by consid-
ering factorial single crosses between biparental populations of pools A and
B. First, elite genotypes A1 to A4 (belonging to pool A) and B1 to B4 (coming
from pool B) will be crossed within each pool for the generation of segregating
inbred populations. In Fig. 7.2, the inbred progenies are represented by genotypes
a* to t* and a to t for pools A and B, correspondingly. Nowadays, inbreds can be
obtained in a singlegeneration by means of doubled haploid techniques (Becker
2011). Thus, eight different populations of full-sib lines are generated, with four
Fig. 7.2 Illustration of a hybrid breeding program using a factorial cross-design between bipa-
rental populations of two pools (namely pools A and B). Genotypes A1 to A4 and B1 to B4 represent
different elite lines belonging to pools A and B, respectively. Circles correspond to the different
families generated by crossing elite lines within each pool (denoted by the symbol), whereas
genotypes a* to t* and a to t are the corresponding progenies from these crosses. Progenies
connected to the same node (family) are assumed as full-siblings. The squares in the center of the
figure represent the 400 possible hybrids obtained between inbreds a* to t* and a to t of pools A
and B, correspondingly. The empty white squares denote hybrids with available phenotypic data,
whereas the shaded ones indicate missing hybrids
populations for each pool. Subsequently, the generated lines are crossed in a
factorial way with genotypes of the opposite pool to evaluate their performance
as parents. For simplicity, only the two extreme levels of relatedness will be
considered as examples. At one extreme, the most related individuals correspond
to hybrids derived from crossing genotypes of a biparental population with the same
genotype of the opposite pool. For instance, this situation is well represented by
hybrids d*b and d*c in Fig. 7.2. At the other extreme, there are less related
hybrids like h*n and q*s, in which the genotypes being crossed do not share any
of the elite lines involved in the generation of biparental populations. Nevertheless,
according to Sect. 7.2.4, in RRS, the best parent lines recognized within each pool
will be subsequently used as founder material of the upcoming segregating inbred

populations during the next breeding cycle. Presumably, it is possible that in some
occasions, two or more of these good parent lines belong to the same biparental
cross (for example, genotypes p and q in Fig. 7.2), which implies that the elite pool
available to serve as parent lines of the different segregating populations would not
always include completely unrelated individuals (Bernardo 1994). Moreover, it is
also anticipated in Fig. 7.2 that the phenotypes of some hybrids will be missed
(represented as shaded squares). This could be because of some evaluation plots
that were missed during the crop season, an insufficient number of seeds available
for field testing, a limited budget for the plant breeding program that ultimately
limited the number of hybrids tested, among other reasons. Furthermore, this
unbalanced scenario provides a very good opportunity to perform GS for the
individuals without phenotypic records. For example, implementing GS to perform
within-population prediction could allow plant breeders to partially testcross a
particular segregating population and then to predict the untested individuals with
a model whose marker effects were estimated using the tested population fraction
(Krchov and Bernardo 2015; Windhausen et al. 2012). This particular breeding
scenario can be found in an illustrative manner in Fig. 7.2 by crossing individuals
from the A1 A2 population with the tester m of the opposite pool, being three of
the five possible hybrids available for marker effects estimation. Later, perfor-
mances of the two remaining hybrids, i.e., b*m and e*m, could be predicted
by GS. Accordingly, within-population prediction schemes have been applied in
GS studies to obtain cross-validated prediction accuracies for testcross performance
of biparental populations in crops like rye (Wang et al. 2014), sugar beet
(Würschum et al. 2013) and maize (Albrecht et al. 2011; Krchov and Bernardo
2015; Lorenzana and Bernardo 2009; Zhao et al. 2012a). Prediction within families
is a closed system and corresponds to the most favorable scenario for GS, which
results in the maximum attainable GS predictabilities (Crossa et al. 2013). This is
mainly because of the combination of high levels of relatedness between estimation
and prediction sets plus the long-range haplotype blocks within families leading to
high linkage disequilibrium (LD) between markers and the loci with true effects on
traits (Albrecht et al. 2011) in addition to the absence of population structure
expected for this situation (Crossa et al. 2013). However, this approach has two
main limitations:
(i) If the estimated marker effects are used for prediction of selection candidates
that are less related to the estimation population, there is a potential risk of drop
in predictability (Albrecht et al. 2014; Habier et al. 2007; Meuwissen 2009;
Meuwissen et al. 2001; Wang et al. 2014).
(ii) Predictability levels in within-population prediction could be constrained by a
limited number of genotypes used for marker effect estimation (Albrecht et al.
2014; Lehermeier et al. 2015; Meuwissen 2009). In this sense, increasing the
size of the estimation set is expected to create more recombination events that
allow an increased resolution for marker effects estimation, ultimately leading
to a model with superior predictability for progenies, which are several gener-
ations away from the estimation set (Lorenz 2013).
Therefore, it has been proposed that a way to obtain robust marker effect
estimates would be to combine the data from different biparental populations in a
single and comprehensive estimation set followed by within-population prediction
(Albrecht et al. 2014; Lehermeier et al. 2015; Wang et al. 2014; Zhao et al. 2012a).
For instance, all the phenotypic data available in Fig. 7.2 could be used to estimate
marker effects, and the remaining missing hybrids would be predicted by
GS. Nevertheless, in some occasions, it seems that marker population interac-
tions (when marker effects are not the same in all populations) could negatively
impact the predictabilities obtained by approaches like this one (Albrecht et al.
2011, 2014; Lehermeier et al. 2015; Zhao et al. 2012a). In general, neither of the
following GS approaches could improve predictabilities compared with a model, in
which marker effects are simply estimated across testcross populations: including a
general population effect, excluding markers with significant marker population
interaction (Zhao et al. 2012a), or modeling population-specific marker effects
considering a variance–covariance structure between populations (Lehermeier
et al. 2015). Presumably, the different levels of relatedness expected in breeding
programs allow keeping acceptable within-population predictability levels when
marker effects are assumed constant across populations, and all these populations
constitute a big combined estimation set. In consequence, this last GS approach
could be a good choice for robust marker effects estimation due to its simplicity.
Cross-Validation Methods Considering Different Levels of Relatedness

in Factorial Crosses
In complete factorial mating designs, a b combinations are possible, with a and

b being the number of lines belonging to pools A and B, respectively (Fig. 7.3,
based on schemes from Schrag et al. 2009). A basic scheme to perform cross-
validation in factorial crosses is the leave-one-out method (Fig. 7.3a). In this cross-
validation scheme, a b – 1 hybrid genotypes are used as the estimation set to
predict the remaining (a b)th genotype. Then, after obtaining genomic predictions
for all the a b hybrids in an iterative manner, these values are compared with the
observed ones (Jacobson et al. 2014). The concept behind this method is the
prediction of a small number of unintentionally missing hybrids that in practice
failed (Schrag et al. 2009). Predictions of testcross performance in maize (Jacobson
et al. 2014) and of sunflower hybrid performance (Reif et al. 2013) correspond to
some examples in which this cross-validation method has been applied. In practice,
however, a large number of early candidate lines of each pool will be only tested as
parents with the best lines of the opposite group because the evaluation of an
extremely large number of hybrids from the a b combination becomes unfeasible
(Schrag et al. 2009). This situation is better represented by the L-shaped cross-
validation scheme (Fig. 7.3b), which has been used as the T2 validation sets in
simulated and experimental data for testcrosses in maize (Technow et al. 2012,
2014) and experimental data for factorial crosses of diversity panels in wheat
(Gowda et al. 2014; Mirdita et al. 2015; Zhao et al. 2015).
Fig. 7.3 Methods of cross-validation for genomic selection (GS) in factorial crosses: (a) Leave-
one-out, (b) L-shaped, (c) Chess-board-like and (d) Mixed scheme. In all cases, two pools of
parent lines (namely pool A and B) were considered. Each square corresponds to a different hybrid
between pools A and B. The empty white squares represent hybrids with available phenotypic data
(estimation set), whereas the shaded ones correspond to different hybrids being predicted by GS
(validation sets). The degree of shading denotes the level of expected relatedness between
estimation and validation sets (the darker the shading, the higher the relatedness), according to
the number of parent lines shared between the hybrids of the estimation and validation sets.
Numbers 0, 1 and 2 indicate none, one or two parents in common between estimation and
validation sets, respectively (based on schemes presented by Schrag et al. 2009)
In the leave-one-out and L-shaped schemes, both parents of the predicted

hybrids in the validation set were already evaluated as parents for other hybrids
in the estimation set (hybrids denoted with the number 2 in Fig. 7.3a, b). Never-
theless, although estimation and validation sets are related through common parents
in these situations, this does not mean that these cross-validation schemes imply
overoptimistic outcomes and should be avoided. This is mainly because both
methods were designed for scenarios in which predictions rely mainly on related-
ness and plant breeders want to profit from it. In contrast, when the objective is the
introduction of new parent lines into the breeding program, the level of relatedness
between the estimation set and the predicted selection candidates is expected to
decrease. In this situation, only one or none of the parents (hybrids represented by
the numbers 1 and 0 in Fig. 7.3c, correspondingly) will be shared between the
estimation set and the predicted hybrids (Schrag et al. 2009). Therefore, LD should
play a more important role than relatedness for the predictability of GS at this point
(Habier et al. 2007). The chess-board-like scheme mimics this scenario (Fig. 7.3c)
and corresponds to the T1 and T0 validation sets used in simulated and experimen-
tal data for factorial crosses in maize (Technow et al. 2012, 2014) and experimental
data for factorial crosses in wheat (Gowda et al. 2014; Mirdita et al. 2015; Zhao
et al. 2015). In addition, a cross-validation scheme with no shared parents between
estimation and validation sets has been used for GS in wheat (Miedaner et al. 2013;
Zhao et al. 2013a, b, 2014a). Last but not least, in reality, plant breeders would use
the same estimation set for all the above-mentioned prediction scenarios; in con-
sequence, a mixed cross-validation scheme (Fig. 7.3d) will be expected in hybrid
breeding programs (Gowda et al. 2014; Mirdita et al. 2015; Technow et al. 2012,
2014; Zhao et al. 2015). Interestingly, prediction accuracy levels for hybrid grain
yield performance of the T2 validation set were similar in maize (Technow et al.
2014) and wheat (Zhao et al. 2015), but even though prediction accuracies
decreased when shifting from T2 to T0 validation sets in both species, this decay
in prediction accuracy was much more pronounced in wheat than in maize. One
possible explanation for this phenomenon is that Zhao et al. (2015) based their
conclusion on a data set concerning factorial crosses of a diversity panel in wheat,
whereas Technow et al. (2014) relied on factorial crosses of maize lines belonging
to a RRS program. As it was already mentioned in section “Relatedness and Its
Implications on the Predictabilities of GS Within Hybrid Breeding Programs”, lines
belonging to a particular pool can be closely related in a RRS program, implying
that there could be some degree of residual relatedness between the estimation and
the T0 validation sets in this particular breeding scheme.
7.3.1.2 Discrepancies Between Test and Target Environments Are

Expected to Impact the Cross-Validated Prediction Accuracy
of GS
Genotype environment interaction is expected to negatively impact predictabil-

ities of GS when the target environments for the selection candidates differ from the
environments considered to test the genotypes used in the estimation set (Albrecht
et al. 2014; Krchov et al. 2015; Schulz-Streeck et al. 2013; Wang et al. 2014;
Windhausen et al. 2012). Furthermore, even though the target locations could be
exactly the same between the estimation set and the selection candidates, the
genotype year interaction can still have a potential negative impact on pre-
dictabilities (Krchov et al. 2015; Wang et al. 2014). Moreover, different
agronomical practices can also derive in different environments. For instance,

phenotypic data coming from field trials with high levels of nitrogen fertilization
could be used as estimation set for genomic predictions of selection candidates
targeted to marginal environments agriculture. In consequence, across-environment
cross-validation schemes have been applied to mimic this situation in studies on GS
for testcross performance in maize (Albrecht et al. 2014; Krchov et al. 2015;
Schulz-Streeck et al. 2013; Windhausen et al. 2012; Zhang et al. 2015) and rye
(Wang et al. 2014). Nevertheless, although these approaches are expected to give
more realistic predictability estimates than cross-validating with the same group of
environments used for the estimation set, most studies on GS for hybrid crops have
ignored this issue (Krchov et al. 2015). Hopefully, future studies would consider
across-environment cross-validation approaches more often, leading to potentially
lower but more realistic predictability values for GS.
7.3.2 Accommodating Dominance Effects Within the GS

Model
Because dominance is a particular feature of hybrid genotypes, the accommodation

of dominance effects within the GS models will receive special attention in this
section.
7.3.2.1 Model Based on Marker Effects
A general model for GS including the dominance component is defined as follows

(Zhao et al. 2013b):
Y ¼ 1n μ þ ZA a þ ZD d þ e, ð7:3Þ
where Y is the n-length vector of phenotypic values pertaining to a particular trait,

1n corresponds to a n-length vector of ones, μ stands for the general mean, whereas
ZA and ZD are n m design matrices for additive and dominance effects of m bi-
allelic markers, respectively. The elements of ZA are coded as 0, 1, 2 according to
the homozygous (first allele), heterozygous and homozygous (second allele) states,
whereas the elements of ZD are 0, 1 for the homozygous and heterozygous states at
the ith locus, correspondingly. The m-length vectors a ¼ (a1, a2, . . .am)T and d ¼
(d1, d2, . . .dm)T contain the elements ai and di, which denote the additive and
dominance effects for the ith marker, respectively, whereas e ¼ (e1, e2, . . .en)T is a
vector of length n and ej is the residual for the jth genotype (Zhao et al. 2013b).
Equation (7.3) could be modified and alternatively expressed in terms of combining
abilities (Piepho 2009).
Ridge Regression Best Linear Unbiased Prediction
An efficient way to obtain the estimate of μ (b μ) along with the predictions for a (^
a)
and d ðdÞ^ of Eq. (7.3) is by means of Ridge Regression-Best Linear Unbiased
Prediction (RR-BLUP) (Whittaker et al. 2000). In this method, assumed that a i
it is
and di marker effects follow the normal distributions N 0; σ 2a and N 0; σ 2d ,
correspondingly, being σ 2a and σ 2d the constant variances of additive and dominance
effects, respectively
(Zhao
et al. 2013b). A normal distribution is assumed for the
residuals ej N 0; σ 2e , where σ 2e is the residual variance of Eq. (7.3). Then, the
solution of the mixed-model equations (Henderson 1984), allowing the obtainment
of μ ^ and d,
b, a, ^ corresponds to:
2 3 2 31 2 3
b
μ n 1Tn ZA 1Tn ZD 1Tn Y
6 7 6 T 7 6 T 7
4 a^ 5 ¼ 4 ZA 1n ZTA ZA þ λA Im ZTA ZD 5 4 ZA Y 5, ð7:4Þ
d^ Z T 1n
D ZTD ZA ZTD ZD þ λ D Im ZTD Y
where Im stands for an identity matrix of size m and the shrinkage parameters λA
along with λD are accordingly defined as the ratios λA ¼ σ 2e =σ 2a and λD ¼ σ 2e =σ 2d
(Meuwissen et al. 2001; Zhao et al. 2013b). The λ terms (λA and λD) prevent over
fitting the model and thus allow the estimation of effects for all markers (Piepho
2009).
Bayesian Approaches
Briefly, Bayesian approaches provide a description of how existing knowledge is

modified by experience. The central concept within Bayesian learning is to com-
bine what is already known about the statistical ensemble before the data are
observed—such knowledge is represented in terms of prior probability distribu-
tions—with the information coming from the data. As a result, a posterior distri-
bution is obtained, from which inferences are made using standard probability
calculus techniques, and the outcomes are interpreted probabilistically (Sorensen
and Gianola 2002). In GS, Bayesian statistics are mainly used to relax some of the
assumptions used within the genomic prediction models (Jannink et al. 2010). In the
pioneering study of Meuwissen et al. (2001), two Bayesian methods were intro-
duced, namely BayesA and BayesB. The linear model at the level of the data is
equal to that used for the RR-BLUP approach in Eq. (7.3), excepting for the
assumptions made for σ 2a and σ 2d (Meuwissen et al. 2001; Zhao et al. 2013b).
BayesA
In RR-BLUP, σ 2a and σ 2d are assumed as common variances for all loci effects, being
this assumption not necessarily realistic for all genetic architectures. BayesA
(Meuwissen et al. 2001; Zhao et al. 2013b) gives solution to this problem because
each of the i loci has its own additive (σ 2ai ) and dominance (σ 2di ) variances. Then, the
prior probability distributions for σ 2gi in BayesA (denoting g the a or d effects in
Eq. (7.3), irrespectively)
correspond to a scaled inverted chi-square distribution in
the way χ 2 vg ; S2g , where vg and S2g are the degrees of freedom and the scale
parameter associated to g, respectively. However, these prior probability distribu-
tions will lead to posteriors which cannot be directly used for estimation, because
they would be conditioned to the unknown value of g (Meuwissen et al. 2001). In
consequence, samples are obtained from full-conditional posterior distributions
using methods like Gibbs sampling. The full-conditional posterior for
~ 2 ~
g corresponds to a normal distribution N Z Gi ðZGi gi þ eÞ=θ i ; σ e =θ i , being
T
~
θ i ¼ Z T Z G þ σ 2 =σ 2 , whereas Z G denotes the ith column of ZA or ZD from
Gi i e g i
Eq. (7.3), irrespectively. For σ 2gi , the full-conditional posterior is a scaled inverted

chi-square distribution denoted as g2i þ vg S2g χ 2 vg þ1 (Zhao et al. 2013b).
BayesB
There could be many loci that do not contribute to the variation on traits with less-
complex genetic architectures (loci with σ 2gi ¼ 0); however, this is not considered by
BayesA (Meuwissen et al. 2001). RR-BLUP and BayesA always fit all markers in
the GS model, even if they truly have zero effects on the trait under study. Although
these markers without true effects are expected to have small predicted effects, they
would add noise to the genomic predictions (Habier et al. 2011). In contrast,
BayesB considers a proportion π of markers whose σ 2gi ¼ 0 and (1 π) with
σ 2gi > 0. Nonetheless, this new consideration makes the usage of Gibbs sampling
unfeasible; hence, the Metropolis-Hastings algorithm has been recommended for
sampling (for a detailed explanation refer to Meuwissen et al. 2001). In addition, in
Sect. 7.3.1.1, it was already mentioned that relatedness between estimation and
validation set influences the prediction ability of GS, and, interestingly, a simula-
tion study showed that BayesB is less impacted by the genetic relatedness among
individuals than RR-BLUP because the former model uses better the information
due to LD (Habier et al. 2007).
BayesCπ
BayesA and BayesB treat π as known, with π ¼ 0 for BayesA and an arbitrary π
value within the range 0 and 1 for BayesB (Habier et al. 2011; Meuwissen et al.
2001), which is in contradiction with the concept of Bayesian learning (Habier et al.
2011; Sorensen and Gianola 2002). To give solution to these drawbacks, Habier
et al. (2011) proposed a new Bayesian approach called BayesCπ, which treats a
common σ 2g and π as unknown with a scaled inverted chi-square and a uniform (0,1)
distribution as prior probability distributions, respectively. An extension for
BayesCπ accommodating dominance effects is described in detail by Zhao et al.

(2013b). Based on results using simulated and animal breeding data, Habier et al.
(2011) recommended BayesCπ for routine applications because of its relatively
short computational time among other advantages.
7.3.2.2 Model Based on Genotypes and Relationship Matrices
Provided that normally the number of loci (or molecular markers) surpasses the
number of genotypes, models based on marker effects are expected to be less
computationally efficient than a model based on genotypic effects (Hayes et al.
2009). By far, the most applied genotype-based method of GS is the one proposed
by VanRaden (2007, 2008). In the literature, this method is often referenced as
GBLUP (Guo et al. 2014; Habier et al. 2011; Su et al. 2012) because of its
similarities with the original best linear unbiased prediction (BLUP) method for
breeding value prediction using pedigree records information presented by Hen-
derson (1984). Considering only additive effects and that the sum ZAa from
Eq. (7.3) corresponds to the breeding value of individuals (VanRaden 2008),
Hayes et al. (2009) demonstrated that GBLUP is equivalent to the mixed models
methods based on markers effects (like RR-BLUP) when the matrix of relationships
among genotypes is calculated from marker profiles. Consequently, there is no
reduction in the prediction accuracy of breeding values by shifting to GBLUP. The
extension of GBLUP for the inclusion of dominance effects is originally defined in
the classical work of Henderson (1985), and it has been lately implemented using
marker-estimated relationship matrices (Da et al. 2014; Su et al. 2012). Hereafter,
we will refer to this method as DGBLUP. In Henderson’s nomenclature, the linear
model underlying DGBLUP looks like Eq. (7.3), but in this case, ZA and ZD are the
design matrices for the a and d vectors of additive and dominance genetic effects of
n genotypes,
correspondingly.
In addition, a and d vectors follow normal distribu-
tions N 0; A∗σ 2A and Nð0, D∗σ 2D Þ, respectively, where σ 2A and σ 2D are now the total
additive and dominance variances, correspondingly. Regarding A and D, they now
correspond to the additive (VanRaden 2007, 2008) and dominance (Da et al. 2014;
Su et al. 2012) marker-estimated relationship matrices, respectively. Then, the
mixed models equations (Henderson 1985) for the DGBLUP are
2 3 2 31 2 3
b
μ n 1Tn ZA 1Tn ZD 1Tn Y
6 7 6 T 7 6 T 7
4 a^ 5 ¼ 4 ZA 1n ZTA ZA þ A1 σ 2e =σ 2A ZTA ZD 5 4 ZA Y 5: ð7:5Þ
d^ ZT 1n
D ZTD ZA ZTD ZD þ D1 σ 2e =σ 2D ZTD Y
Subsequently, the variance component estimates along with the predictions of

genetic effects in Eq. (7.5) are simultaneously computed by the restricted maximum
likelihood (REML) algorithm. Moreover, based on relationship matrices, the mixed
models for hybrid prediction can be modified to accommodate random terms such
as the GCA effects of parents and their corresponding SCAs. For instance, Bernardo
(1994, 1996) used a relationship matrix termed S for the SCA component, which
was expressed as the direct product between the relationship matrices pertaining to
the GCAs of two heterotic groups (Stuber and Cockerham 1966). In recent years,
the GCA plus SCA model using marker-estimated relationship matrices has been
implemented in the context of hybrid genomic prediction (Massman et al. 2013;
Piepho 2009).
7.3.2.3 Classical Mixed Models or Bayesian Approaches?
There are no GS methods that are suitable for all genetic architectures and/or
breeding schemes. Therefore, the superior performance in terms of predictability
of different GS approaches relies always on the context of their applications.
Interestingly, most studies on GS for hybrid genotypes have relied on classical
mixed model predictions by means of GBLUP (Albrecht et al. 2011, 2014; Guo
et al. 2013; Massman et al. 2013; Riedelsheimer et al. 2012; Technow et al. 2014;
Zhao et al. 2015) and RR-BLUP (Gowda et al. 2014; Guo et al. 2013; Hofheinz
et al. 2012; Jacobson et al. 2014; Lorenzana and Bernardo 2009; Massman et al.
2013; Miedaner et al. 2013; Mirdita et al. 2015; Riedelsheimer et al. 2012;
Windhausen et al. 2012; Würschum et al. 2013; Zhao et al. 2012a, b, 2013a, b,
2014a). In contrast, lesser studies have applied Bayesian approaches for hybrid
performance prediction (Lorenzana and Bernardo 2009; Miedaner et al. 2013;
Mirdita et al. 2015; Technow et al. 2014; Zhao et al. 2013a,b, 2014a, 2015). One
reason for these observations could be that the understanding and implementation
of classical mixed model methods is much more straightforward than for Bayes
approaches, which is also facilitated by the large number of user-friendly REML
and BLUP packages available (Guo et al. 2014). However, as it was already
mentioned in section “BayesB”, Bayesian methods like BayesB, which assign
effects equal to zero to a proportion π of markers, are expected to be less impacted
by the relatedness between estimation and validation sets than methods conferring
nonzero effects to all markers available, like RR-BLUP. This particular issue is not
trivial if one takes into consideration that the information from genetic relationships
is halved with each additional generation and that LD information is more persis-
tent through time (Habier et al. 2007). Nevertheless, in situations in which pedigree
relatedness can be efficiently exploited by plant breeders, RR-BLUP could be
valuable for predicting hybrid performance (Zhao et al. 2013b, see Sect. 7.3.1.1
for a detailed explanation). Moreover, in practice, the joint evidence of studies on
hybrid performance prediction has been inconclusive about the superior predict-
ability of Bayesian over classical mixed models approaches (Lorenzana and
Bernardo 2009; Miedaner et al. 2013; Mirdita et al. 2015; Technow et al. 2014;
Zhao et al. 2013a, b, 2014a, 2015). For instance, although it was expected that
predictions obtained by means of BayesCπ outperformed RR-BLUP predictions for
the medium-complexity trait Fusarium head blight resistance in hybrid wheat, both
methods performed equally (Mirdita et al. 2015). The joint evidence suggests, in
consequence, that both groups of methods are in practice relatively equivalent, and
also that GS methods must be ultimately selected based on their implementability
and understandability, which makes methods like GBLUP and RR-BLUP the
preferred ones.
7.3.2.4 Benefits of Modeling Dominance Effects
Studies based on simulated data have shown that the higher the relative importance
of dominance versus additive effects is, the more beneficial (in terms of predict-
ability) the inclusion of dominance over additive effects within the GS models
would be (Guo et al. 2013; Nishio and Satoh 2014; Technow et al. 2012; Zhao et al.
2013b). Nevertheless, evaluating the benefits of including, in general, any effect
within the GS model using empirical data is challenging because often it is not
exactly known which of the following situations is being confronted when no
benefits are observed: a) some assumptions of the GS methods that include the
particular effect evaluated are disrupted; thus, the methods cannot accurately
capture the true effect; b) the GS methods can accurately model the effect under
evaluation, but the influence of the true effect is extremely low and c) the GS
methods cannot accurately capture the effect evaluated, and the influence of the true
effect is also negligible. A study on genomic predictions for grain yield in a
population of 1604 wheat hybrids found some predictability improvements by
using DGBLUP over GBLUP (Zhao et al. 2015), but these benefits were not
observed by means of RR-BLUP and Bayesian approaches in a population
compromising 90 wheat hybrids (Zhao et al. 2013b). In addition, genomic pre-
dictions considering additive plus dominance effects were not superior to predicting
frost tolerance exclusively by additive effects in hybrid wheat, presumably because
of the low contribution of dominance compared with additive effects for this trait
(Zhao et al. 2013a). However, Guo et al. (2013) observed using experimental data
for different traits in an F1 maize population that, in general, the benefits of
including dominance over additive effects were more pronounced when the differ-
ences between broad-sense and narrow-sense heritabilities for the traits were
higher. They expressed broad-sense heritabilities as the ratio of the additive plus
dominance variances estimates to the total phenotypic variance, whereas in the
narrow-sense heritability, only the additive variance was considered as the numer-
ator. Therefore, their findings suggest that the more different were these two values,
the higher was the importance of the dominance variance and, in consequence, the
higher the benefits from including dominance over additive effects. Moreover, in
general, accommodating dominance over additive effects has been also beneficial
in GS for plant height and heading date in a hybrid population of wheat (Zhao et al.
2014a). In conclusion, joint evidence of simulated and experimental data studies
points out that modeling dominance over additive effects is beneficial when dom-
inance effects have an important contribution to the total genetic variation
(Bernardo 1994; Gowda et al. 2013; Guo et al. 2013; Nishio and Satoh 2014;
Reif et al. 2013; Technow et al. 2012; Zhao et al. 2013a, b, 2014a, 2015).
Estimation of genetic parameters such as variance components could give some

insights for the relative contribution of dominance effects to total genetic variance,
which also highlights the importance of phenotypic analyses as a decision tool
before performing GS.
7.3.3 Beyond the Modeling of Dominance
7.3.3.1 Accommodating Epistasis
Using simulated data, Guo et al. (2013) showed that epistasis can bias the pre-
dictions achieved by GS models based solely on additive and dominance main
effects. In the past, models including epistasis were presumably avoided because of
their high computational burden, especially if a large number of markers was
available (Jiang and Reif 2015). Nevertheless, this limitation has also encouraged
scientists to search for more efficient GS methods accommodating epistasis. In
principle, GBLUP can be extended for the inclusion of any order of epistatic
interactions by approximating the epistatic genomic relationship matrix of the
interaction effects with the Hadamard product operation (denoted as #) between
the relationship matrices of main effects (Henderson 1984; Jiang and Reif 2015; Su
et al. 2012). For instance, additive additive, additive dominance and additive
additive dominance interactions are represented as A # A, A # D and A # A # D,
respectively. Hereafter, this method will be called EGBLUP (Jiang and Reif 2015).
Another approach for GS considering epistatic interactions corresponds to the semi-
parametric reproducing kernel Hilbert space (RKHS) regression method (Gianola
et al. 2006; Gianola and van Kaam 2008). Recently, it has been demonstrated that
both RKHS and EGBLUP considering epistasis are similar approaches and, as a
result, reach comparable levels of predictability (Jiang and Reif 2015). Solely
taking into account the additive effects and their interactions, the RKHS method
is at first sight similar to the additive effects GBLUP (VanRaden 2007, 2008), but a
K matrix is used instead of the original A matrix. The K matrix is a n n kernel
matrix whose entries are functions of marker profiles of pairs of genotypes in the
way K ¼ (k(xi, xj)), where k ( ) represents a particular function (e.g., the Gaussian
kernel function), whereas xi and xj are the rows of the marker profile matrix
pertaining to genotypes i and j, correspondingly (Jiang and Reif 2015). Further-
more, Bayesian approaches have been also to accommodate epistatic interactions,
like the empirical Bayes (e-Bayes) method, in which marker additive main effects
and second-order epistatic interactions are calculated based on estimates of true
marker variances (Lorenzana and Bernardo 2009). However, only few studies have
used GS models taking into account epistatic interactions for hybrid performance
prediction. Lorenzana and Bernardo (2009) observed that ignoring epistatic inter-
actions within the GS model leads to higher predictabilities than accommodating
epistasis by means of e-Bayes for different traits in maize testcross populations. In
addition, two recent crop plant studies considered additive, dominance and their
second-level interactions in GS for hybrid performance by means of EGBLUP

(Xu et al. 2014; Zhao et al. 2015). Xu et al. (2014) concluded that in general, adding
dominance plus epistasis over the main additive component did not help to improve
genomic predictions for hybrid performance in rice, whereas Zhao et al. (2015) only
observed some benefits from including the dominance over the additive component,
but there was no further accuracy improvement from including epistasis in hybrid
wheat. Moreover, both authors partly attributed this issue to the fact that the A and
D matrices are correlated with the relationship matrices of epistatic effects; hence,
the two former matrices already capture much of the variation for hybrid prediction.
Nonetheless, Xu et al. (2014) observed by means of simulated data that for large
estimation sets, there is a benefit in prediction by including the epistatic effects
within the model, reflecting the need of large population sizes to accurately take
advantage of epistasis prediction in GS. In consequence, more studies on hybrid
performance prediction are needed to explore the benefits and limitations of GS
approaches, which accommodate epistatic effects.
7.3.3.2 Other GS Approaches
W-BLUP Method
Recently, a new GS method, named weighted best linear unbiased prediction

(W-BLUP), was designed to properly incorporate the information of previously
known functional markers (Zhao et al. 2014a). Alternatively, Bernardo (2014)
suggested modeling known functional markers as fixed effects. In the study of
Zhao et al. (2014a), it was observed that the predictability values for heading date
obtained by means of marker assisted selection (MAS) using the functional marker
Ppd-D1 were higher than by performing GS based on 1280 single nucleotide
polymorphism (SNP) markers in a hybrid wheat population. Nevertheless, when
both types of information were combined using W-BLUP, predictability values
surpassed the ones obtained by MAS or GS alone. In consequence, W-BLUP holds
the promise to bridge the gap between MAS and GS when known functional
markers are available.
Multiple-Trait GS
Simulation studies have shown that prediction accuracies for a trait with relatively
low heritability can be improved when a genetically correlated trait with higher
heritability is included within a multiple-trait GS model (Guo et al. 2014; Hayashi
and Iwata 2013; Jia and Jannink 2012). However, plant studies exploiting these
benefits are scarce, and, to the best of our knowledge, there are only a couple of
studies evaluating the advantages and limitations of these methods for hybrid
prediction. In a study concerning two testcross populations of rye (Schulthess
et al. 2016), grain protein content predictions markedly benefited from the
availability of grain yield information in one of the two testcross populations. These
benefits were even more pronounced when a few phenotypic records were available
for the predicted target trait, but many more phenotypic records were at hand for the
indicator trait during the estimation of marker effects. In addition, Lehermeier et al.
(2015) performed genomic predictions in ten different testcross populations of
maize by means of a multiple-trait GS approach called MG-GBLUP, which con-
sidered each testcross population as if it were a different trait. Nevertheless,
MG-GBLUP was, in general, less than or equivalently accurate to estimating
marker effects by means of a much simpler GS model that assumes the same
marker effect for all testcross populations. In the future, more studies are needed
to evaluate in detail the routine implementation of multiple-trait GS approaches for
hybrid breeding.
Metabolomic Prediction
In the omics era, metabolomics corresponds to the systematic study of metabolite

profiles pertaining to a particular process at the organism, tissue or cell level.
Metabolomics provides a tool for measuring biochemical activity directly by
monitoring the substrates and products transformed during metabolism (Patti
et al. 2012). Nowadays, the availability of massive and automated analytical
platforms has facilitated the routine generation of this high dimensional data
(Patti et al. 2012; Ward et al. 2015). The weak correspondence between the
information of genetic and metabolic profiles obtained from leaves of maize
(Riedelsheimer et al. 2012) and wheat (Zhao et al. 2015) suggests that both
information sources content connected but, at the same time, different biological
information (Riedelsheimer et al. 2012). In this sense, it is expected that metabolite
profiles condense genetic and environmental influences together (Feher et al. 2014).
The basic model of metabolomic prediction is similar to Eq. (7.3) but omitting the
dominance term and respecifying the design matrix for a, corresponding this last
term now to the vector of metabolite effects. Accordingly, the new Z matrix for
metabolite effects contains the normalized metabolite levels instead of the 1, 0, 1
nomenclature originally used for additive effects of bi-allelic markers
(Riedelsheimer et al. 2012). However, it should be noted that metabolite profiles
from parental lines interact in a very complex way to determine the metabolite
profiles of hybrids. For example, a particular metabolite found at low levels in
parents A, B and C can be at high and low levels in the hybrids A B and B C,
respectively. This could be a consequence of variation in a second metabolite at the
average parental levels of crosses A B and B C, implying that metabolite levels
cannot be regarded as independent from each other. In addition, the influence of
dominance between parental metabolite profiles can make the situation even more
complex (Feher et al. 2014). Riedelsheimer et al. (2012) performed metabolomic
prediction for different traits in maize testcrosses by using 130 leaf metabolite
profiles and reported prediction accuracy levels that were only slightly lower than
those achieved based on highly dense genomic profiles of 38,019 SNP markers.
Nevertheless, the results of this comparison should be carefully interpreted, because

accuracies of metabolomic prediction were previously normalized by the repeat-
ability of metabolite profiles, leading to an overestimation of the predictabilities
achieved by metabolomic prediction. Lately, Zhao et al. (2015) found that pre-
dictions for grain yield based on 34 leaf-extracted metabolomic profiles reached
substantially inferior prediction accuracies compared with GS based on 17,372 SNP
markers in hybrid wheat. Furthermore, Riedelsheimer et al. (2012) and Zhao et al.
(2015) observed no benefits in terms of prediction accuracies from combining
genomic and metabolomic information into a single prediction model. In the future,
further studies on prediction by means of models that simultaneously integrate
genomic, transcriptomic, proteomic, and metabolomic information would be
needed; thus, helping to understand how these different layers of biological infor-
mation interact to shape the complex phenotypes of hybrid plants.
Considering Marker Environment Interactions
It was already mentioned in Sect. 7.3.1.2 that genotype environment interactions

have the potential to negatively impact the predictability of GS. Considering this
issue and by means of single-stage RR-BLUP approaches, Schulz-Streeck et al.
(2013) partitioned the additive marker effects of Eq. (7.3) into main and marker
environment interaction effects for testcross performance prediction in maize. In a
first attempt, the authors predicted untested genotypes in tested environments (i.e.,
environments already included within the estimation set), and they observed that
GS predictabilities improved when shifting from a model with only additive effects
to a model including main additive plus marker environment interaction effects.
Moreover, similar results were in general obtained by Zhang et al. (2015) for
testcross performance prediction of grain yield considering well-watered and
water-stressed environments, although the advantages of accommodating genotype
environment interactions within the GS model were less pronounced for days to
anthesis and plant height prediction. Nevertheless, Schulz-Streeck et al. (2013) also
found that these benefits disappeared when predictions were performed for untested
genotypes in untested environments, highlighting the importance of across-
environment cross-validation schemes to evaluate the prospects of GS in a more
realistic manner. More studies should be conducted in this research area to elucidate
if GS models including marker environment interactions or approaches that
dissect these interactions by considering environmental covariates (Heslot et al.
2014) have promising applications in hybrid crop species.
7.3.4 Implementation of GS in Hybrid Breeding
So far, the reader has probably realized that most studies on GS for hybrid
performance in plants have mainly focused on the factors influencing the predict-
ability of GS. For sure, this vast amount of knowledge has helped researchers and
the plant breeding community in how to carefully interpret predictability values
according to different biological and breeding scenarios and also in how to improve
it when possible. Lately, studies have started to extensively discuss the implemen-
tation of GS in plant breeding. Because prediction within families corresponds to
the most favorable scenario for the implementation of GS (Crossa et al. 2013),
implementation studies have mainly focused on plant breeding programs based on
biparental populations (Endelman et al. 2014; Krchov and Bernardo 2015; Longin
et al. 2015; Lorenz 2013; Riedelsheimer and Melchinger 2013). In general, the
authors concluded that GS should be considered as a tool to assist plant breeding in
a similar way as MAS; hence, it is not intended to completely replace phenotypic
selection (PS). Therefore, as already stated in section “Relatedness and Its Impli-
cations on the Predictabilities of GS Within Hybrid Breeding Programs”, the main
idea would be to partially testcross a particular biparental population and then to
predict the remaining individuals using a model previously trained with the tested
population fraction (Krchov and Bernardo 2015; Windhausen et al. 2012). Here lies
the paradigm shift of GS because the purpose of phenotypic evaluations turns from
exclusively guiding PS toward additionally calibrating statistical models for GS
(Lorenz 2013). Subsequently, selection is supposed to be simultaneously performed
in, both, estimation and prediction sets (Endelman et al. 2014; Krchov and
Bernardo 2015; Riedelsheimer and Melchinger 2013). Certainly, one general ques-
tion that arises in studies on GS implementation is: Given a limited budget, what is
the optimal allocation of resources between estimation and prediction sets that
maximizes the ratio of selection gains per unit of time between GS and pure PS? In
principle, the essence of this problem can be represented by means of a mathemat-
ical model, in which a particular objective function, which is subjected to some
constraints, aims to be maximized based on the optimization of some decision
variables according to certain parameters. Moreover, this dissected representation
can potentially clarify the interrelationships between the different components of
the problem, thus, facilitating its comprehension and analysis (Hillier and
Lieberman 2001). In fact, GS implementation has been already described as a
nonlinear optimization problem (Endelman et al. 2014; Riedelsheimer and
Melchinger 2013). Relying on the available literature along with our own concepts
and criteria about the topic, the current section of this chapter aims to introduce the
problem of GS implementation for a testcrossed biparental population pertaining to
a hybrid breeding program.
7.3.4.1 Towards a Successful Implementation of GS in Hybrid

Breeding
Objective Function: The Relative Efficiency of GS over Pure PS
The objective function is the particular value aimed to be maximized

(or minimized) throughout the optimization process and should be one of the first
model components to be identified or defined. This value measures the performance
expressed as a mathematical function of the decision variables (Hillier and
Lieberman 2001). For instance, plant breeders rely on the selection gain (ΔG) to
compare and measure the performance of different selection methods. In principle,
ΔG is the difference of the mean genetic value between the offspring of the selected
fraction and the whole population before selection. This value can be predicted as:
ΔG ¼ ihσ G , ð7:6Þ
being i the selection intensity, whereas h and σ G denote the square roots of the
heritability (h2) and the genetic variance of the population (σ 2G), correspondingly. In
truncation selection (one-tail selection), the i term is a function of the proportion of
individuals being selected according to a particular threshold located away from the
mean phenotype of the original population before selection (Falconer and Mackay
1996). Moreover, h2 is expressed as follows:
σ 2G
h2 ¼ σ 2GE
, ð7:7Þ
σ2
σ 2G þ Nr: Env þ Nr: EnvNr:
e
Rep
where σ 2GE and σ 2e correspond to the genotype environment interaction and

residual variance components, respectively, whereas Nr. Env and Nr. Rep are the
number of environments and replicates used in balanced field tests, correspondingly
(Holland et al. 2003; Piepho and M€ ohring 2007). Then, h is interpreted as the
accuracy of PS (Endelman et al. 2014; Lorenz 2013; Riedelsheimer and Melchinger
2013), whereas the accuracy of GS can be mechanistically decomposed as:
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
λh2
GS accuracy ¼ , ð7:8Þ
λh2 þ 1
being λ ¼ MNE
e
, where NE is the number of genotyped and phenotyped individuals in
the estimation set and Me is the effective number of loci (Daetwyler et al. 2008).
Nevertheless, Eq. (7.8) is presented here only for explanatory purposes because the
relationship between NE and GS accuracy should be ultimately determined in an
empirical manner based on previous available genomic and phenotypic data from
plant breeding institutions (Endelman et al. 2014; Krchov and Bernardo 2015). On
top of this, considering that selection would be simultaneously performed within,
both, estimation and prediction sets, the overall ΔGGS of plant breeding programs
based on GS should be a function of the ΔG achieved by GS for the genomic
predicted fraction in addition to the ΔG within the estimation set (Endelman et al.
2014; Krchov and Bernardo 2015; Riedelsheimer and Melchinger 2013). In this
sense, because genomic predictions can also be obtained for the genotypes
conforming the estimation set at no extra costs, these predictions could also be
combined with the phenotypic data in a molecular selection index (Endelman et al.
2014; Lorenz 2013; Riedelsheimer and Melchinger 2013) as originally proposed for
MAS more than two decades ago by Lande and Thompson (1990). Therefore,
accuracies of selection within the estimation and prediction sets will be differen-
tially denoted as rE and rP, correspondingly. Consequently and according to
Eq. (7.6), ΔGGS can be expressed as:

SE ð S SE Þ
ΔGGS ¼ iE r E þ iP r P σ G , ð7:9Þ
S S
where S and SE are the number of genotypes selected from the whole population and
the estimation set, respectively, whereas iE and iP are the selection intensities within
the estimation and prediction sets, correspondingly (Endelman et al. 2014;
Riedelsheimer and Melchinger 2013). A similar formulation to Eq. (7.9) was
presented by Krchov and Bernardo (2015). Ultimately, the relative efficiency
(RE) of GS compared with pure PS should be maximized. This metric is defined
as the ratio between ΔGGS and the previously maximized ΔG for pure PS given the
same assumptions or conditions. In this sense, ΔGGS should be compared with the
best ΔG attainable by means of pure PS in the same biparental population. Thus, the
breakeven point for a successful GS implementation is when the RE of GS reaches
unity because values above this threshold reflect a better competitiveness for GS
over pure PS (Endelman et al. 2014; Krchov and Bernardo 2015; Riedelsheimer and
Melchinger 2013). Last but not least, because GS has the potential to accelerate
plant breeding programs by reaching more selection cycles than pure PS within the
same amount of time, more important than the direct comparison between ΔGGS
and ΔG of pure PS is the contrast between their ΔG per unit of time (Longin et al.
2015). Examples of GS implementation with more than one stage of selection can
be found elsewhere (Endelman et al. 2014; Longin et al. 2015; Lorenz 2013);
however and for simplicity, we will approach the optimization problem by solely
considering one-stage selection.
The Main Constraint: The Budget
The constraints are any restrictions on the values that the decision variables can
take, and they are mathematically expressed by means of inequalities or equations
(Hillier and Lieberman 2001). In other words, restrictions basically shape the space
of all possible (optimal and suboptimal) solutions for the optimization problem.
In consequence, the basic constraint for any plant breeding strategy would be that
its costs do not exceed the available budget. In the context of GS, this is well
represented by the following inequation:
N E CE þ ðN N E ÞCP Budget, ð7:10Þ
where CE and CP correspond the costs of estimation and prediction sets, respec-
tively, whereas N is the total number of individuals in the biparental population
before selection. Costs and budget can be conveniently expressed in plot equiva-
lents or yield plot units, that is, the cost of phenotyping one yield plot (Endelman
et al. 2014; Riedelsheimer and Melchinger 2013). Moreover, costs can be further
decomposed as follows: CP corresponds to the costs of producing a line (CL) plus
the costs of genotyping (CG), whereas CE ¼ CL + CG + CF, being CF the cost of the
field trials using a number of plots determined by the Nr. Env Nr. Rep combina-
tion. In this sense, Eq. (7.10) represents the trade-off existent in GS between the
number of plots used for field evaluations (phenotyping intensity) and NE (Lorenz
2013). In parallel, the constraint for pure PS is reduced to: N(CL + CF) Budget,
reflecting that the trade-off between the number of individuals and phenotyping
intensity is also present in pure PS (Endelman et al. 2014). Similar formulations of
costs can be found in studies on GS implementation (Endelman et al. 2014; Krchov
and Bernardo 2015; Longin et al. 2015; Lorenz 2013; Riedelsheimer and
Melchinger 2013).
So far, studies on GS implementation have considered, either explicitly or
indirectly, that the quantity of seed pertaining to the F1 of the testcross will be
enough to perform sufficient field trials. However, the production of F1 seeds is a
well-known constraint in hybrid breeding, especially in naturally self-pollinated
species (Longin et al. 2012; Whitford et al. 2013). Plant breeding institutions give
solution to this problem, for instance, by increasing the planting area for the parent
line(s), which in turn results in higher costs (Longin et al. 2012). Nevertheless, GS
could allow the prediction of testcross performance for parent lines, which do not
produce enough F1 seeds for testing in field trials. Being this an advantage for GS
over PS, future studies on GS implementation should consider a seed quantity
constraint to explore this potential benefit of GS.
Decision Variables
Decision variables are a quantitative representation of the decision to be made.

They conform the group of all quantities that will be changed (optimized) along the
space of solutions (determined by the constraints) during the optimization process
and, ultimately, leading to the maximization (or minimization) of the objective
function at their optimum values (Hillier and Lieberman 2001). According to the
objective function and constraint presented, the decision variables for the optimal
allocation of resources in GS correspond to:
(i) NE: provided N was previously optimized for pure PS, optimizing NE would
elucidate which proportion of the population should be, both, phenotyped and
genotyped, or only genotyped to maximize the RE of GS considering a
particular budget and other assumptions.
(ii) iE and iP: given N and NE along with preknown properties of the probability
distributions for the testcross performances of estimation and prediction sets,
the optimization of iE and iP would determine the fraction of top-ranking
genotypes that should be selected within each respective set (Burrows 1975;
Falconer and Mackay 1996). Ultimately, this optimized fraction would set the
optimum values for SE and (S-SE), indicating the number of individuals that
should be selected from estimation and prediction sets, respectively, and
allowing the maximization of the RE of GS according to the particular budget
and other assumptions.
(iii) Nr. Env and Nr. Rep: the optimization of the Nr. Env Nr. Rep combination
will indicate the phenotyping intensity within the estimation set, which is
necessary to maximize the RE of GS taking on count the particular budget
limitation and other assumptions.
Parameters
Parameters are the constants (coefficients and right-hand sides) present in the
constraints as well as in the objective function (Hillier and Lieberman 2001) and,
in contrast to the decision variables, they are assumed as fixed known values. The
main parameters considered in the present formulation of GS implementation
problem are the budget and costs (CL, CG and CF). However, because of the
uncertainty associated to the actual values of parameters, assigning the proper
quantities to them is a very delicate task. In consequence, it would be important
to investigate how the solution for the optimization problem would change if other
possible values were assigned to the parameters. Accordingly, sensitive parameters
will correspond to all those constants whose value cannot be modified without
changing the optimal solution of the problem. Thus, special care should be allo-
cated to the assignment of values for those particular parameters (Hillier and
Lieberman 2001). For instance, even though it has been shown that there is a
great amount of flexibility in choosing the optimal NE and phenotyping intensity
levels that maximize ΔGGS (Lorenz 2013; Riedelsheimer and Melchinger 2013),
this flexibility for the optimal solution strongly relies on the available budget
because smaller budgets will restrict the set of possible solutions that maximize
ΔGGS in one-stage selection (Riedelsheimer and Melchinger 2013). In addition, the
budget and CG will determine whether or not GS can compete with pure PS
(RE 1), with higher budgets (Krchov and Bernardo 2015; Riedelsheimer and
Melchinger 2013) and lower CG (Endelman et al. 2014; Krchov and Bernardo 2015;
Riedelsheimer and Melchinger 2013) having in general positive effects on the RE
of GS. Furthermore, the influence of CG on RE of GS becomes more important with
smaller budgets (Riedelsheimer and Melchinger 2013). Therefore, studies on GS
implementation often speculate about a future decline of CG, implying a more

favorable scenario for GS implementation (Endelman et al. 2014; Krchov and
Bernardo 2015; Riedelsheimer and Melchinger 2013).
Last but not least, Me, σ 2G , σ 2GE and σ 2e are estimated values according to a
particular population, test environments and target trait. Like any estimator, they
would rely on the sample used for estimation, and hence, their values would be also
subjected to uncertainty. Nevertheless, the impacts of changes in these parameters
on the decision-making process in GS have been less studied. Briefly, high values of
Me would reflect the polygenic nature of a particular trait (Falconer and Mackay
1996). Assuming different Me values for grain yield testcross performance in maize,
Riedelsheimer and Melchinger (2013) observed that Me has a negative influence on
the RE of GS, being this trend more pronounced when low budgets are available.
Additionally, these authors also studied the influence of the importance of σ 2GE and
σ 2e relative to σ 2G (σ 2GE : σ 2G and σ 2e : σ 2G ratios, respectively) on the solutions for the
optimal allocation of resources in GS. Under the assumption of a high budget, and
that CG is less than the cost of phenotyping one plot, Riedelsheimer and Melchinger
(2013) showed that the RE of GS stays nearly constant at 1.3 regardless of the
assumed value for σ 2GE : σ 2G and also, that increasing the σ 2e : σ 2G ratio has a positive
effect on the RE of GS. The authors attributed their observations to the differences
in reallocation of resources between GS and pure PS that concomitantly occurred
when varying σ 2GE : σ 2G and σ 2e : σ 2G . Nonetheless, future studies should evaluate if
these last observations hold truth for a less favorable scenario with more severe
budget limitations and current genotyping costs.
7.3.4.2 Model Recalibration After a Successful GS Implementation

in Early Breeding Stages: A Proposal
It was already highlighted in section “The Main Constraint: The Budget” that the
decisions of plant breeders are always restricted by a limited budget; hence, they
will always confront a trade-off between phenotyping intensity and the number of
individuals being tested in field trials. Moreover, from section “The Main Con-
straint: The Budget”, it is also concluded that even ignoring budget limitations, the
limited quantity of seed belonging to the F1 of a testcross would be an additional
constraint for the phenotyping intensity of the testcross performance, especially
during early breeding stages. In addition, it becomes clear from Eq. (7.7) that a
restriction in phenotyping intensity is expected to decrease the potentially achiev-
able h2, and, according to Eq. (7.8), this limitation in h2 will constrain the GS
accuracy at any given NE. Nevertheless, as a plant breeding program proceeds, the
phenotyping intensity will be increased for the individuals being selected. Thus,
incorporating this high-quality phenotypic data into the estimation set has the
potential to improve prediction accuracies for future lines from (or very close
related to) the biparental population in which GS was originally implemented.
Consequently, new marker effects will be obtained by using updated estimation
sets, allowing the recalibration of GS models. At the same time, new phenotypic
data will be available for those genotypes whose early selection relied only on
genomic predictions and this new information could be used to increase the size of
the estimation set, which in turn would further improve GS accuracy according to
Eq. (7.8). However, there are two important points that should be taken on count
before proceeding with the GS recalibration:
First, there is an intrinsic issue of selection in plant breeding which has being
ignored within the present proposal for GS recalibration: plant breeders want to
increase the frequency of favorable alleles within the selected set (Bernardo 2010),
leading to a decrease in σ 2G and to the misrepresentation of alleles with negative
effects for a particular trait in the selected fraction. In this sense, updating the
estimation set with the new information would mostly increase the phenotyping
intensity for favorable alleles and, at the same time, would increase the frequency of
favorable alleles within the estimation set. The impacts of PS within the estimation
set on GS accuracy were studied by Zhao et al. (2012b) using grain yield testcross
performance data of maize. They observed that using one-tail selection combined
with high selection intensities within the estimation set led to a substantial decrease
in GS accuracy compared with unselected estimation sets of the same size. In
addition, they found that unselected estimation sets with low phenotyping intensi-
ties reached higher GS accuracies than estimation sets with much higher
phenotyping intensity but subjected to one-tail selection and high selection inten-
sity. Interestingly, their results also showed that estimation sets subjected to
bidirectional selection (two-tails selection) in combination with high selection
intensities reached superior GS accuracies than unselected estimation sets of the
same size. Moreover, they concluded that just a small proportion of low performing
genotypes (10–15% of the total fraction selected from both tails) would be enough
for this purpose. Nevertheless, updating the estimation set by means of this last
strategy implies a paradigm shift of selection within plant breeding programs and,
consequently, should be further analyzed under the eye of the “economics” of GS
for its wide acceptance by the plant breeding community.
Second, a simulation study showed that GS predictability does not always
coincide with the accuracies at the individual level (Clark et al. 2012), implying
that even though GS could reach high predictability or accuracy levels (according
to the definitions in Sect. 7.3.1), some genotypes within the predicted set would be
very accurately predicted and others would not. Prediction accuracy at the genotype
level is often termed as “reliability,” and it has been extensively used in the field of
animal science (Clark et al. 2012; Mrode 2005; VanRaden 2008) and later in the
context of GS in hybrid crops (Akdemir et al. 2015; Rincent et al. 2012). Further-
more, the simulation study of Clark et al. (2012) also showed that the reliability is
highly associated to the maximum level of relatedness between estimation set and
the particular genotype being predicted. In other words, highly reliable predictions
are expected for genotypes, which were very well represented by a few or even by a
single extremely closely related genotype(s) within the estimation set. Additionally,
the reliability criterion has been used to identify individuals that are best suited for
the conformation of the estimation set within diverse populations of maize

(Akdemir et al. 2015; Rincent et al. 2012). In this sense, both studies in maize
demonstrated that, in general, estimation sets that maximize the average reliability
of the prediction set lead to higher GS predictabilities or accuracies than estimation
sets constructed by random sampling. Nonetheless, the impacts of using the reli-
ability of predicted genotypes as threshold criteria for updating the information
within the estimation set of biparental populations have not been evaluated so far;
therefore, studies will be needed to elucidate if reliabilities are also useful to further
improve or maintain accuracies in the context of GS recalibration.
7.3.4.3 Final Words for the Implementation of GS in Hybrid Breeding
GS implementation in hybrid breeding is challenging because of the number of

variables and imponderables involved in the optimal allocation of resources
between estimation and prediction sets that influence the RE of GS under a given
budget constraint. The current section was intended to give readers the basic
conceptual framework of this problem considering biparental populations. It is
anticipated that the problem formulation presented here is a simplification of the
real picture; hence, readers are encouraged to make use of available decision
support software (e.g., Endelman et al. 2014; Riedelsheimer and Melchinger
2013) to get further insights into how decision variables and parameters influence
the RE of GS. Furthermore, readers with intermediate to advanced programming/
planning skills are invited to elaborate their own models to find an optimal solution
for the GS implementation problem. Preferably, the GS implementation should be
planned and evaluated in an integrated manner and from a multidisciplinary point
of view, considering together the skills and knowledge of plant breeders, scientists,
technicians, process engineers and managers. In the future, once several plant
breeding institutions and companies have already implemented GS, people should
start to analyze successful and unfortunate cases of study to gain the empirical
knowledge that is necessary to bridge the gap between theory and practice for the
GS implementation problem in hybrid breeding.
References
Akdemir D, Sanchez JI, Jannink JL (2015) Optimization of genomic selection training populations
with a genetic algorithm. Genet Sel Evol 47:38
Albrecht T, Wimmer V, Auinger HJ, Erbe M, Knaak C, Ouzunova M, Simianer H, Sch€ on CC
(2011) Genome-based prediction of testcross values in maize. Theor Appl Genet 123:339–350
Albrecht T, Auinger HJ, Wimmer V, Ogutu JO, Knaak C, Ouzunova M, Piepho HP, Sch€ on CC
(2014) Genome-based prediction of maize hybrid performance across genetic groups, testers,
locations, and years. Theor Appl Genet 127:1375–1386
Becker H (2011) Pflanzenzüchtung (in German). Auflagennr. 2. Verlag Eugen Ulmer, Stuttgart
Bernardo R (1994) Prediction of maize single-cross performance using RFLPs and information
from related hybrids. Crop Sci 34:20–25
Bernardo R (1996) Best linear unbiased prediction of maize single-cross performance. Crop Sci
36:50–56
Bernardo R (2010) Breeding for quantitative traits in plants. Stemma Press, Woodbury
Bernardo R (2014) Genomewide selection when major genes are known. Crop Sci 54:68–75
Bos I, Caligari P (2008) Selection methods in plant breeding, 2nd edn. Springer, Dordrecht
Bruce AB (1910) The Mendelian theory of heredity and the augmentation of vigor. Science
32:627–628
Burrows PM (1975) Expected selection differentials for directional selection. Biometrics
28:1091–1100
Clark SA, Hickey JM, Daetwyler HD, Van der Werf JHJ (2012) The importance of information on
relatives for the prediction of genomic breeding values and implications for the makeup of
reference populations in livestock breeding schemes. Genet Sel Evol 44:4
Collins GN (1921) Dominance and vigor of first generation hybrids. Am Nat 55:116–133
Comstock RE, Robinson HF, Harvey PH (1949) A breeding procedure designed to make maxi-
mum use of both general and specific combining ability. Agron J 41:360–367
Crossa J, Pérez-Rodrı́guez P, Hickey J, Burgue~
no J, Ornella L, Cerón-Rojas J et al (2013) Genomic
prediction in CIMMYT maize and wheat breeding programs. Heredity 112:48–60
Crow JF (1948) Alternative hypotheses of hybrid vigor. Genetics 33:477–487
Da Y, Wang C, Wang S, Hu G (2014) Mixed model methods for genomic prediction and variance
component estimation of additive and dominance effects using SNP markers. PLoS One 9:
e87666
Desta ZA, Ortiz R (2014) Genomic selection: genome-wide prediction in plant improvement.
Trends Plant Sci 19:592–601
East EM (1936) Heterosis. Genetics 21:375–397
Endelman JB, Atlin GN, Beyene Y, Semagn K, Zhang X, Sorrells ME, Jannink JL (2014) Optimal
design of preliminary yield trials with genome-wide markers. Crop Sci 54:48–59
Falconer DS, Mackay TFC (1996) Introduction to quantitative genetics, 4th edn. Ronald Press
Company, New York
Feher K, Lisec J, R€ omisch-Margl L, Selbig J, Gierl A, Piepho HP, Nikiloski Z, Willmitzer L
(2014) Deducing hybrid performance from parental metabolic profiles of young primary roots
of maize by using a multivariate Diallel approach. PLoS One 9:e85435
Gianola D, van Kaam JB (2008) Reproducing kernel Hilbert spaces regression methods for
genomic assisted prediction of quantitative traits. Genetics 178:2289–2303
Gianola D, Fernando RL, Stella A (2006) Genomic-assisted prediction of genetic value with
semiparametric procedures. Genetics 173:1761–1776
Gowda M, Zhao Y, Maurer HP, Weissmann EA, Würschum T, Reif JC (2013) Best linear unbiased
prediction of triticale hybrid performance. Euphytica 191:223–230
Gowda M, Zhao Y, Würschum T, Longin CFH, Miedaner T, Ebmeyer E, Schachschneider R,
Kazman E, Schacht J, Martinant JP, Mette MF, Reif JC (2014) Relatedness severely impacts
accuracy of marker-assisted selection for disease resistance in hybrid wheat. Heredity
112:552–561
Guo T, Li H, Yan J, Tang J, Li J, Zhang Z, Zhang L, Wang J (2013) Performance prediction of F1
hybrids between recombinant inbred lines derived from two elite maize inbred lines. Theor
Appl Genet 126:189–201
Guo G, Zhao F, Wang Y, Zhang Y, Du L, Su G (2014) Comparison of single-trait and multiple-
trait genomic prediction models. BMC Genet 15:30
Habier D, Fernando RL, Dekkers JCM (2007) The impact of genetic relationship information on
genome-assisted breeding values. Genetics 177:2389–2397
Habier D, Fernando R, Kizilkaya K, Garrick D (2011) Extension of the bayesian alphabet for
genomic selection. BMC Bioinf 12:186
Hallauer AR, Carena MJ, Miranda Filho JB (2010) Quantitative genetics in maize breeding. Iowa
State University Press, Ames
Hayashi T, Iwata H (2013) A Bayesian method and its variational approximation for prediction of
genomic breeding values in multiple traits. BMC Bioinf 14:34
Hayes BJ, Visscher PM, Goddard ME (2009) Increased accuracy of artificial selection by using the
realized relationship matrix. Genet Res 91:47–60
Henderson CR (1984) Applications of linear models in animal breeding. University of Guelph,
Guelph
Henderson CR (1985) Best linear unbiased prediction of non-additive genetic merits. J Anim Sci
60:111–117
Heslot N, Akdemir D, Sorrells M, Jannink JL (2014) Integrating environmental covariates and
crop modeling into the genomic selection framework to predict genotype by environment
interactions. Theor Appl Genet 127:463–480
Hillier FS, Lieberman GJ (2001) Introduction to operations research, 7nd edn. McGraw Hill,
New York
Hjorth JSU (1994) Computer intensive statistical methods. Validation model selection and boot-
strap. Chapman & Hall, London
Hofheinz N, Borchardt D, Weissleder K, Frisch M (2012) Genome-based prediction of test cross
performance in two subsequent breeding cycles. Theor Appl Genet 125:1639–1645
Holland JB, Nyquist WE, Cervantes-Martińex CT (2003) Estimating and interpreting heritability
for plant breeding: an update. In: Janick J (ed) Plant breeding reviews, vol 22. Wiley,
New York, pp 9–112
Hull FH (1945) Recurrent selection for specific combining ability in corn. J Am Soc Agron
37:134–145
Jacobson A, Lian L, Zhong S, Bernardo R (2014) General combining ability model for
genomewide selection in a biparental cross. Crop Sci 54:895–905
Jannink JL, Lorenz AJ, Iwata H (2010) Genomic selection in plant breeding: from theory to
practice. Brief Funct Genomics 9:166–177
tion accuracy. Genetics 192:1513–1522
Jiang Y, Reif JC (2015) Modeling epistasis in genomic selection. Genetics 201:759–768
Jones DF (1917) Dominance of linked factors as a means of accounting for heterosis. Genetics
2:466–479
Keeble F, Pellew C (1910) The mode of inheritance of stature and of time of flowering in peas
(Pisum sativum). J Genet 1:47–56
Krchov LM, Bernardo R (2015) Relative efficiency of genomewide selection for testcross perfor-
mance of doubled haploid lines in a maize breeding program. Crop Sci 55:2091–2099
Krchov LM, Gordillo GA, Bernardo R (2015) Multienvironment validation of the effectiveness of
phenotypic and genomewide selection within biparental maize populations. Crop Sci
55:1068–1075
Lande R, Thompson R (1990) Efficiency of marker-assisted selection in the improvement of
quantitative traits. Genetics 124:743–756
Lehermeier C, Sch€on CC, de los Campos G (2015) Assessment of genetic heterogeneity in
structured plant populations using multivariate whole-genome regression models. Genetics.
doi:10.1534/genetics.115.177394
Longin CFH, Mühleisen J, Maurer HP, Zhang H, Gowda M, Reif JC (2012) Hybrid breeding in
autogamous cereals. Theor Appl Genet 125:1087–1096
Genet 128:1297–1306
Lorenz AJ (2013) Resource allocation for maximizing prediction accuracy and genetic gain of
genomic selection in plant breeding: a simulation experiment. G3 3:481–491
selection in biparental plant populations. Theor Appl Genet 120:151–161
Massman JM, Gordillo A, Lorenzana RE, Bernardo R (2013) Genomewide predictions from maize
single-cross data. Theor Appl Genet 126:13–22
Melchinger AE, Gumber RK (1998) Overview of heterosis and heterotic groups in agronomic
crops. In: Lamkey KR, Staub JE (eds) Concepts and breeding of heterosis in crop plants.
ASACSSA-SSSA Publication, Madison, pp 29–44
Meuwissen THE (2009) Accuracy of breeding values of ‘unrelated’ individuals predicted by dense
SNP genotyping. Genet Sel Evol 41:35
Miedaner T, Zhao Y, Gowda M, Longin CFH, Korzun V, Ebmeyer E, Kazman E, Reif JC (2013)
Genetic architecture of resistance to Septoria Tritici blotch in European wheat. BMC Geno-
mics 14:858
Mirdita V, Liu G, Zhao Miedaner T, Longin CFH, Gowda M, Mette MF, Reif JC (2015) Genetic
architecture is more complex for resistance to Septoria Tritici blotch than to Fusarium head
blight in central European winter wheat. BMC Genet 16:430
Mrode RA (2005) Linear models for the prediction of animal breeding values, 2nd edn. CABI
Publishing, Wallingford
Nishio M, Satoh M (2014) Including dominance effects in the genomic BLUP method for genomic
evaluation. PLoS One 9:e85792
Patti GJ, Yanes O, Siuzdak G (2012) Innovation: metabolomics: the apogee of the omics trilogy.
Nat Rev Mol Cell Biol 13:263–269
Piepho HP (2009) Ridge regression and extensions for genomewide selection in maize. Crop Sci
49:1165–1176
Piepho HP, M€ohring J (2007) Computing heritability and selection response from unbalanced
plant breeding trials. Genetics 177:1881–1888
Reif JC, Gumpert F, Fischer S, Melchinger AE (2007) Impact of genetic divergence on additive
and dominance variance in hybrid populations. Genetics 176:1931–1934
Reif JC, Zhao YS, Würschum T, Gowda M, Hahn V (2013) Genomic prediction of sunflower
hybrid performance. Plant Breed 132:107–114
Richey FD (1942) Mock-dominance and hybrid vigor. Science 96:280–281
Riedelsheimer C, Melchinger AE (2013) Optimizing the allocation of resources for genomic
selection in one breeding cycle. Theor Appl Genet 126:2835–2848
Riedelsheimer C, Czedik-Eysenberg A, Grieder C, Lisec J, Technow F, Sulpice R, Altmann T,
Stitt M, Willmitzer L, Melchinger AE (2012) Genomic and metabolic prediction of complex
heterotic traits in hybrid maize. Nat Genet 44:217–220
Rincent R, Laloë D, Nicolas S, Altmann T, Brunel D, Revilla P, Rodriguez VM, Moreno-
Gonzalez J, Melchinger A, Bauer E (2012) Maximizing the reliability of genomic selection
by optimizing the calibration set of reference individuals: comparison of methods in two
diverse groups of maize in breds (Zea mays L.) Genetics 192:715–728
Schnell FW, Cockerham CC (1992) Multiplicative vs. Arbitrary gene action in heterosis. Genetics
131:461–469
Schrag TA, Frisch M, Dhillon BS, Melchinger AE (2009) Marker-based prediction of hybrid
performance in maize single-crosses involving doubled haploids. Maydica 54:353–362
Schulthess AW, Wang Y, Miedaner T, Wilde T, Reif JC, Zhao Y (2016) Multiple-trait- and
selection indices-genomic predictions for grain yield and protein content in rye for feeding
purposes. Theor Appl Genet 129:273–287
Schulz-Streeck T, Ogutu JO, Gordillo A, Karaman Z, Knaak C, Piepho HP (2013) Genomic
selection allowing for marker-by-environment interaction. Plant Breed 132:532–538
Sorensen D, Gianola D (2002) Likelihood, Bayesian, and MCMC methods in quantitative genet-
ics. Springer, New York
Stuber CW, Cockerham CC (1966) Gene effects and variances in hybrid populations. Genetics
54:1279–1286
Su G, Christensen OF, Ostersen T et al (2012) Estimating additive and non-additive genetic
variances and predicting genetic merits using genome-wide dense single nucleotide polymor-
phism markers. PLoS One 7:e45293
Technow F, Riedelsheimer C, Ta S, Melchinger AE (2012) Genomic prediction of hybrid
performance in maize with models incorporating dominance and population specific marker
effects. Theor Appl Genet 125:1181–1194
Technow F, Schrag TA, Schipprack W, Bauer E, Simianer H, Melchinger AE (2014) Genome
properties and prospects of genomic prediction of hybrid performance in a breeding program of
maize. Genetics 197:1343–1355
Tracy WF, Chandler MA (2006) The historical and biological basis of the concept of heterotic
patterns in corn belt dent maize. In: Lamkey KR, Lee M (eds) Plant breeding: the Arnel R
Hallauer international symposium. Blackwell Publishing, Ames, pp 219–233
VanRaden PM (2007) Genomic measures of relationship and inbreeding. Interbull Bull 37:33–36
VanRaden PM (2008) Efficient methods to compute genomic predictions. J Dairy Sci
91:4414–4423
Wang Y, Mette MF, Miedaner T, Gottwald M, Wilde P, Reif JC, Zhao Y (2014) The accuracy of
prediction of genomic selection in elite hybrid rye populations surpasses the accuracy of
marker-assisted selection and is equally augmented by multiple field evaluation locations
and test years. BMC Genomics 15:556
Ward J, Rakszegi M, Bedo Z, Shewry P, Mackay I (2015) Differentially penalized regression to
predict agronomic traits from metabolites and markers in wheat. BMC Genet 16:19
Whitford R, Fleury D, Reif JC, Garcia M, Okada T, Korzun V, Langridge P (2013) Hybrid
breeding in wheat: technologies to improve hybrid wheat seed production. J Exp Bot
64:5411–5428
Genet Res 75:249–252
Windhausen VS, Atlin GN, Hickey JM, Crossa J, Jannink JL, Sorrels ME, Raman B, Cairns JE,
Tarekegne A, Semagn K, Beyene Y, Grudloyma P, Technow F, Riedelsheimer C, Melchinger
AE (2012) Effectiveness of genomic prediction of maize hybrid performance in different
breeding populations and environments. G3 2:1427–1436
Wricke G, Weber WE (1986) Quantitative genetics and selection in plant breeding. Gruyter, Berlin
Würschum T, Reif JC, Kraft T, Janssen G, Zhao Y (2013) Genomic selection in sugar beet
breeding populations. BMC Genet 14:85
Xu S, Zhu D, Zhang Q (2014) Predicting hybrid performance in rice using genomic best linear
unbiased prediction. Proc Natl Acad Sci U S A 111:12456–12461
Zhang X, Pérez-Rodrı́guez P, Semagn K, Beyene Y, Babu R, López-Cruz MA, San Vicente F,
Olsen M, Buckler E, Jannink JL, Prasanna BM, Crossa J (2015) Genomic prediction in
biparental tropical maize populations in water-stressed and well-watered environments using
low-density and GBS SNPs. Heredity 114:291–299
Zhao Y, Gowda M, Liu W, Würschum T, Maurer HP, Longin CFH, Ranc N, Reif JC (2012a)
Accuracy of genomic selection in European maize elite breeding populations. Theor Appl
Genet 124:769–776
Zhao Y, Gowda M, Longin CFH, Würschum T, Ranc N, Reif JC (2012b) Impact of selective
genotyping in the training population on accuracy and bias of genomic selection. Theor Appl
Genet 125:707–713
Zhao Y, Gowda M, Würschum T, Longin CFH, Korzun V, Kollers S, Schachschneider R, Zeng J,
Fernando R, Dubcovsky J (2013a) Dissecting the genetic architecture of frost tolerance in
Central European winter wheat. J Exp Bot 64:4453–4460
Zhao Y, Zeng J, Fernando R, Reif JC (2013b) Genomic prediction of hybrid wheat performance.
Crop Sci 53:802–810
Zhao Y, Mette MF, Gowda M, Longin CFH, Reif JC (2014a) Bridging the gap between marker-
assisted and genomic selection of heading time and plant height in hybrid wheat. Heredity
112:638–645
Zhao Y, Mette MF, Reif JC (2014b) Genomic selection in hybrid breeding. Plant Breed. doi:10.
1111/pbr.12231
Zhao Y, Li Z, Liu G, Jiang Y, Maurer HP, Würschum T, Mock HP, Matros A, Ebmeyer E,
Schachschneider R, Kazman E, Schacht J, Gowda M, Longin CFH, Reif JC (2015) Genome-
based establishment of a high-yielding heterotic pattern for hybrid wheat breeding. Proc Natl
Acad Sci U S A. doi:10.1073/pnas.1514547112
Chapter 8
Opportunities and Challenges to Implementing
Genomic Selection in Clonally Propagated
Crops
Dorcus C. Gemenet and Awais Khan
8.1 Clonally Propagated Crops
Clonal, or vegetative, crops are asexually propagated, that is, successive mitoses of
specialized plant tissues develop into a new clonal population from a single mother
plant (Bisognin 2011). Asexual propagation is used in all important root and tuber
crops, many forage crops, almost 75% of perennial fruit trees, wooden ornamentals,
many cut flowers, pot plants, and forest trees (Miller and Gross 2011; Denis and
Bouvet 2013; Grunenberg et al. 2009), and presents a number of advantages. It can
lead to increased levels of heterozygosity, fix favorable combinations of important
traits, eliminate undesirable crosses and the resulting deleterious effects, and allow
easy identification and propagation of favorable mutations. It is also an efficient
method for maintenance, conservation, and in vitro and ex vitro propagation of
cultivars with no viable seeds (Bisognin 2011). Despite these benefits, breeding of
clonally propagated crops also includes several challenges. Most clonal fruit and
forest trees have long juvenile phases, extensive outcrossing, widespread hybridi-
zation, limited population structure, multiple origins, and ongoing crop–wild gene
flow, and have suffered from mild domestication bottlenecks due to clonal propa-
gation (Miller and Gross 2011). Many clonally propagated crops are polyploid,
which enables them to adapt to rapidly changing environment by maintaining
increased heterozygosity, thus reducing inbreeding depression (Griffin et al.
2011). The high natural heterozygosity means that these crops are not amenable
D.C. Gemenet
International Potato Centre (CIP), Apartado 1558, Lima 12, Peru
A. Khan (*)
Plant Pathology and Plant-Microbe Biology Section, School of Integrative Plant Science,
Cornell University, New York State Agricultural Experiment Station, Geneva, NY 14456,
USA
e-mail: awais.khan@cornell.edu

DOI 10.1007/978-3-319-63170-7_8
186 D.C. Gemenet and A. Khan
to self-pollination due to high inbreeding depression. These effects lead to

decreased second- and third-order favorable interactions and a reduced frequency
of trigenic and tetragenic loci interactions, as well as the possibility to accumulate
masked deleterious recessive alleles (Bradshaw 1994). Overdominance and epi-
static interactions at a given locus of a heterozygous genotype can mask deleterious
effects, but they can emerge after selfing. Genetic studies in polyploid crops are
often further complicated by the presence of different ploidy levels. Cultivated
potato (Solanum sp.) could be diploid (2n ¼ 2x), triploid, (2n ¼ 3x), tetraploid,
(2n ¼ 4x), or pentaploid (2n ¼ 5x) (Spooner et al. 2010); cultivated sweet potato
(Ipomoea batatas) is hexaploid (2n ¼ 6x) (Jones 1965); and the ploidy level of
sugarcane is yet unknown (Gouy et al. 2013), whereas Musa species are triploid
(2n ¼ 3x) (Simmonds 1962). In these crops, genetic analysis is complicated by the
presence of multiple alleles at a given locus, mixed inheritance patterns, association
between ploidy and mating system variation, among others (Dufresne et al. 2014).
The presence of several segregating alleles at each locus of highly heterozygous
clonally propagated crops make their breeding challenging. For example, in potato,
an autotetraploid, four alleles per loci could be segregating, meaning that crossing
two tetra-allelic potatoes could result in 32 genetically distinct genotypes with
different levels of trait expression, different from the original parents. Therefore,
potato breeders need to evaluate large populations to find at least one genotype with
the desirable allelic combinations (Jansky et al. 2016). Expanding to the six
different alleles possible for autohexaploid sweet potato, the number of genotypes
to be screened will be far too large to find desirable trait combinations from
multiple different loci. In this chapter, we discuss conventional breeding methods
and their challenges, the potential of genomic selection (GS), challenges for its
implementation with some examples, and outlook of GS as a population improve-
ment strategy in clonal crops.
8.2 Breeding Strategies for Clonal Crops
According to Simmonds (1979), breeding clonal crops requires a crossing step to

provide sexual seeds, as a break from the normal clonal propagation, and to create
genetic variation that can be exploited during selection in subsequent cycles, before
reverting to clonal selection. This crossing step involves two heterozygous parents
to produce clonally propagated hybrids and is a distinctive feature in these crops.
The resulting hybrids are therefore heterozygous and heterogeneous, and can
display all forms of genetic effects (additive, dominance, overdominance, and
epistatic; Ceballos et al. 2015). The cross is followed by phenotypic mass selection
or recurrent selection. These conventional approaches are both time and resource
consuming, as they involve crossing in one generation, planting of true seed plants
in another generation, and observation of clones from selected true seed plants over
several generations and different environments, to evaluate genotype-by-environ-
ment (G x E) interactions (Gruneberg et al. 2009). Another challenge in breeding,
8 Opportunities and Challenges to Implementing Genomic Selection in Clonally. . . 187
selection, and conducting yield trials for clonally propagated crops is producing
enough planting material. Due to their heterozygosity and self-incompatibility,
selfing may result in inbreeding depression, and crossing to make seed can take a
long time. Therefore, vegetative propagules are produced and then used for field
and selection trials. However, each plant can only make a fixed number of vegeta-
tive propagules, and multiplication can thus be slow and costly. For example, it
takes about 45 years from a cross to have enough planting material for replicated
multienvironment field trials in cassava (Ceballos et al. 2015), whereas in tree fruits
the breeding cycle may extend to a dozen years or more (van Nocker and Gardiner
2014). Another challenge, in several clonally propagated crops, is maintaining
disease-free (“clean” “good seed quality,” or low disease) stocks (clones) during
the clonal-increase phase; otherwise, trait evaluations could be significantly
affected. Genome-based selection, especially at the seedling stage of true seed
plants, can be an important approach towards expediting the breeding process by
shortening the lengthy selection cycle, removing the need of several subsequent
clonal selection cycles and reducing the time required for multiplication (Myles
2013). Resources are also saved by maintaining fewer genotypes for phenotypic
evaluation. Genome-based selection can be achieved through using either one of
two methods: Firstly, conventional marker-assisted selection (MAS) using diag-
nostic markers linked with a few qualitative and quantitative trait loci (QTL) with
large effects; secondly, GS using genomic estimated breeding values (GEBVs)
predicted by high-density genome-wide molecular markers to select superior prog-
eny (Meuwissen 2007) or combining both (Spindel et al. 2015). Conventional
selection (phenotypic selection, PS) is usually based on multiple traits. In both PS
and GS, a multistep selection and/or indexes could be developed to filter and select
clones with the best combinations of traits of interest (Fig. 8.1).
8.3 Key Features of Genomic Selection
Genomic selection (GS) is a method of selection proposed by Meuwissen et al.

(2001) using genome-wide genotypic data to predict the phenotypic performance of
a genotype by estimating its breeding value or total genetic value, referred to as
GEBV. In the GS process, statistical models are used to estimate the relationship
between phenotypes and genotypes in a subset of the population normally referred
to as a training population. The models developed are then tested with a validation
set, which is a subset of the population phenotyped in the same environment(s) as
the training set, a process called cross-validation. Validated models are then applied
to a breeding population with only genotypic data to determine GEBV of the
genotypes, and finally, elite individuals with desirable trait combinations are
selected based on these GEBVs only (Nakaya and Isobe 2012). Several statistical
models have been proposed for GS, each with different assumptions on marker
effects and the relationship between the markers. However, for a given predictive
model to perform well for a given trait, it has to follow the continuum of the genetic
Fig. 8.1 (a) A conventional breeding scheme for clonally propagated crop includes crossing,
selection, and yield trials. (b) A genomic-assisted scheme for clonally propagated crop includes
crossing, selection, and yield trials
architecture of that trait (Poland and Rutkoski 2016). For quantitative traits, mixed
models such as genomic best linear unbiased prediction (G-BLUP) and ridge-
regression best linear unbiased prediction (RR-BLUP) are widely used, as they
mimic the conventional best linear unbiased prediction (BLUP) normally used in
phenotypic selection (PS). Several Bayesian models work better for traits that fall
in-between quantitative and qualitative, that is, they are regulated by a few major
genes. These include BayesA, BayesB (Meuwissen et al. 2001), BayesCπ (Habier
et al. 2011), and Bayessian LASSO (least absolute shrinkage and selection operator;
Park and Casella 2008). However, all these methods assume additive genetic
effects, which is not the case for clonally propagated crops. For clonal crops,
these methods would be adequate if GS was only applied to select for new parents
for the crossing step, in which case, additive genetic variation would be important.
However, for variety development in clonal crops, both additive and nonadditive
genetic effects are important. Methods that model both additive and nonadditive
effects would then be required if GS was to be successfully deployed for variety
development in these crops (Azevedo et al. 2015). For these crops, models such as
reproducing kernel Hilbert space (RKHS; Gianola and van Kaam 2008) and random
forest (RF; Breiman 2001), which have been shown to capture both additive and
nonadditive effects, could be applied. The predictive ability for each model is
estimated as the correlation between the observed breeding values and the predicted
breeding values from GS, whereas the prediction accuracy is calculated by dividing
the predictive ability by the mean heritability of the validation set (Poland and
Rutkoski 2016). As such, this prediction accuracy is affected by several factors.
These include the size of the training set: the larger the training set, the better the
accuracy of prediction; the heritability of the trait: the higher the heritability of the
trait, the better the accuracy; relationship between training and breeding
populations: the closer the relationship, the better the accuracy; linkage disequilib-
rium (LD) between the marker and the trait: the higher the LD, the better the
prediction accuracy. The LD between markers and QTLs (traits) also determines
the number of markers required to perform GS: plants with rapid LD decay, like
outcrossing species, require more markers, whereas less markers may be required
for inbreeding species (Spindel et al. 2015; Nakaya and Isobe 2012). Other factors
reported to affect prediction accuracy include population structure (pedigree),
environment, data redundancy, epistasis, type of cross-validation (i.e. fivefold,
tenfold, Jacknife), GS prediction models and accuracy calculation approach. GS
requires a clear definition of the breeding scenario in which selection will be
implemented and a detailed analysis of the population structure. Larger training
sets that are closely related to the target breeding and selection populations give
higher prediction accuracy. GS studies in maize have shown that a major portion of
the prediction accuracy estimated using prediction models developed with
unrelated populations comes from population structure and is affected by environ-
ment. Higher prediction accuracy can be achieved by also modeling GE (Genotype
Environment) and borrowing information from related environments (Crossa
et al. 2014; Windhausen et al. 2012).
8.4 Challenges to Implementing Genomic Selection

in Clonal Crops
8.4.1 Modeling of Genetic Effects and Heritability
The unique features of the population and quantitative-genetic parameters of

clonally propagated crops may pose challenges to the adoption of GS models
currently developed for seed-propagated crops. Most of the proposed models in
GS mainly model additive effects and assume dominance and epistatic effects as
part of the residual. This holds true for seed-propagated crops, as it is not possible to
transfer nonadditive genetic effects to the next generation sexually, rather new
nonadditive combinations are formed during each sexual recombination cycle.
However, for clonally propagated crops, dominance and epistatic effects play an
important role beside additive effects and need special consideration. This is
because the whole set of alleles, together with their interactions, are passed to the
next generation through clonal propagation. Ceballos et al. (2015) demonstrated the
presence of additive, dominance, and epistatic effects in cassava whose magnitudes
differed for individual traits. Because gene action is locus and trait specific, the
currently available GS models will give different prediction accuracy for different
traits in clonally propagated crops. In sugarcane, Gouy et al. (2013) reported similar
predictive accuracy for several GS models on a single trait but significantly

different predictive accuracy for each model on different traits. This means that
additive, dominance, and epistatic genetic effects must be properly analyzed for
every trait to select the best model for each. Most often, specialized populations
with specific mating designs are needed to estimate the extent of these gene actions,
and so far, these populations have not been properly defined for most traits of
importance in clonally propagated crops. Furthermore, estimation of narrow sense
heritability, one of the key factors affecting prediction accuracy in GS (Nakaya and
Isobe 2012), is mainly based on additive genetic variation, which holds true for
seed-propagated crops, but not for clonal crops. This then calls for the use of
additive, dominance, and epistatic genetic effects in calculating broad sense heri-
tability estimates used in GS for clonally propagated crops. Munoz et al. (2014)
used both pedigree-based and marker-based information to model additive, domi-
nance, and first-order epistatic interaction effects in the tree species Pinus taeda.
They concluded that prediction accuracy of GEBV improved by including additive
and nonadditive effects to the predictive models. Wolfe et al. (2016) used both
additive only and additive plus nonadditive effect models to show that including
nonadditive effects in the model improved prediction accuracy. On the other hand,
including large effect QTL as fixed effects in additive-only model improved
prediction accuracy for cassava mosaic disease resistance. Prediction accuracies
ranged from 0.53 to 0.58 with different models, indicating that GS would be useful
for selecting cassava mosaic disease resistance.
8.4.2 Linkage Disequilibrium between Markers

and Quantitative Trait Loci
Linkage disequilibrium (LD), the nonrandom association of alleles at different loci,

is another factor affecting prediction accuracy in GS (Nakaya and Isobe 2012). As
opposed to linkage, which refers to the physical connection of loci on a chromo-
some and are inherited together, LD refers to the correlation among alleles in the
whole population (Flint-Garcia et al. 2003). LD breaks down both by recombination
(intrachromosomal LD) and independent assortment (interchromosomal LD), as
well as by other factors that affect the Hardy–Weinberg equilibrium (Flint-Garcia
et al. 2003). Estimation of LD in clonally propagated crops is only possible during
the true seed plant stage resulting from meiotic events at the crossing step. The
heterozygosity and heterogeneous nature of most clonal species ensures a large
breakdown of LD at this crossing step. Because GS assumes that the marker density
used is large enough that all genes are in LD with some of the markers (Meuwissen
2007), this implies that several markers are required to ensure higher prediction
accuracy of the GEBV in clonally propagated crops relative to seed-propagated in
which selfing may be possible. There has been little systematic evaluation of the
extent of LD in most clonally propagated crops; in most cases, the effective number
of markers required for efficient GS is unknown. Even for those crops where LD has
been studied a little, there are contrasting findings regarding LD depending on the
marker systems used. In potato, Simko et al. (2006), using SNPs within 100 bp
derived from bacterial artificial chromosome (BAC) ends, reported LD extending to
10 cM possibly due to the shorter physical distance of the SNPs, whereas D’hoop
et al. (2010) reported LD extending to 5 cM using over 3000 AFLP markers,
whereas Stich et al. (2013) reported LD decay after 275 bp using genome-wide
SNPs from the SOLCAP array. In sugarcane, Jannoo et al. (1999) reported LD
extending to 10 cM, whereas Raboin et al. (2008) reported LD ranging from
0–30 cM. In case of cassava, LD analysis based on SNPs from GBS showed
decay between 10–50 kb (Wolfe et al. 2016). In banana (Musa sp), Sardos et al.
(2016) showed that LD extended to 10–100 kb. In apple, Kumar et al. (2013)
reported LD persisting to approximately 1 cM. For crops like sweet potato and
yam, efforts are still underway at the International Potato Center (CIP) and the
International Institute of Tropical Agriculture (IITA). Findings from the above
studies, with the exception of the findings by Stich et al. (2013), indicate that LD
persists longer for polyploid clonal crops compared with diploid clonal crops. This
can be attributed to the bottleneck in breeding polyploid crops using only few
parents and the confounding effects of polyploidy on marker identification. Despite
this, LD in clonal crops in general persists longer compared with outcrossing seed-
propagated crops. In maize, Yan et al. (2009) showed LD decay within 1–10 kb on a
global maize collection. This aspect can be attributed to clonal propagation that
ensures reduced meiotic events. It is important to precisely estimate LD for most of
these crops if GS were to be successful as this will enhance determining effective
population sizes and genotyping densities that have great impact on the accuracy of
genomic prediction (Grattapaglia and Resende 2011).
8.4.3 Genetic Architecture of Traits and Size of Training

Population
The genetic architecture of a trait of interest affects prediction accuracy of GS

(Nakaya and Isobe 2012). Genetic architecture of a trait is a complex of factors,
including the number of genes controlling the trait and their genomic location, the
effects of substituting alleles of these genes and the heritability of a trait (Poland
and Rutkoski 2016). Many clonally propagated crops are self-incompatible and
polyploid, resulting in multiple alleles and dosages at a given locus (Slater et al.
2016). Therefore, the allele combinations responsible for a given trait are numerous
and mostly unknown. To apply GS in a breeding program, the training population
used to develop prediction models should be large enough to capture representative
combinations of alleles for traits of interest in the breeding population (Jannink
et al. 2010). This is important because selection reduces additive genetic variation
and reduces genetic gains in subsequent generations, whereas development of
commercial varieties requires simultaneous improvement of many quantitative

traits (Poland and Rutkoski 2016). For clonal crops, there has been little genetic
gain for complex traits like yield compared with cereals, possibly due to using only
a few parents and reduced meiotic combinations due to clonal propagation (Slater
et al. 2016). So, to avoid bottlenecks from breeding, the effective population sizes
may need to increase with increasing ploidy levels. Effective population size refers
to the number of individuals who contribute offspring to the next generation while
meeting the Hardy–Weinberg equilibrium. However, increasing effective popula-
tion size also leads to more rapid decay of LD and reduced prediction accuracy
(Grattapaglia and Resende 2011). No systematic analysis of the effective popula-
tion sizes required for accurate estimation of GEBV has been studied in most
clonally propagated crops. Initial estimation of effective population size for tetra-
ploid potato did indicate the need for 79 initial parents (Slater et al. 2016).
Furthermore, traits controlled by several small effect loci across the genome with
complex genetic architecture are more responsive to genotype-by-environment
interaction (G x E). Ly et al. (2013), using 17 traits in cassava, showed that
prediction accuracy reduced between 0.01 to 0.18 for different locations compared
with the training set, indicating a strong relationship between prediction accuracy
and G x E. Resende et al. (2012) also showed reducing prediction accuracies with
geographical distance between the model development sites and validation sites in
loblolly pine (Pinus taeda). Because cassava and loblolly pine are diploid clonally
propagated crops, it could be speculated that prediction accuracy for higher ploidy
levels will be further reduced for traits showing strong G E interactions. There-
fore, the initial investment in development of GS models may be higher for clonally
propagated crops compared with seed-propagated crops because of the need to
evaluate larger numbers of training sets across several target environments, where
environment here refers to both sites and seasons.
8.4.4 Number of Generations Following Training Model
One of the main attractions to implementing GS for crop improvement programs is

the potential of lower costs and shorter generation intervals, arising from the ability
to predict GEBVs with high accuracy, early in the breeding cycle, over several
generations, without phenotyping selection populations. Prediction accuracy of GS
depends on LD between markers and QTL, and is expected to decline in the
generations following the population initially used for developing the training
model for the estimation of GEBVs (Habier et al. 2007; Nakaya and Isobe 2012).
The composition and genetic distance of individuals in a training population also
affects prediction accuracy and differs for traits and the heritability of traits (Weng
et al. 2016a, b). This is expected to be much more important in seed-propagated
crops, where new combinations of nonadditive genetic effects are formed during
each sexually reproductive cycle (Grattapaglia and Resende 2011). This may be
less of a problem in clonally propagated crops, especially if the training population
is sufficiently related to the breeding population because there is only one recom-
bination step (the crossing step) per selection cycle, which can recur, and fixed
genetic effects are fully passed on to the next generation through clones. Ly et al.
(2013) did not find significant reduction in prediction accuracy due to relatedness or
lack of relatedness between training and prediction sets in cassava. So, although
initial investment for developing GS models may be higher for clonally propagated
crops due to the higher marker density required to cover for a relatively large
number of effective population size, the cost may be evened out by application of
the same models for longer than for seed-propagated crops.
8.5 Examples of Genomic Selection in Clonal Crops
Most of the GS carried out in clonally propagated crops thus far has been proof-of-
concept, that is, developing GS models on training sets, estimating prediction
accuracies, and testing models in validation sets. Extension and implementation
into practical breeding and selection programs has not been achieved fully, espe-
cially in the public sector. Here, we enumerate a few examples where GS has been
tested in clonally propagated crops. de Oliveira et al. (2012) used the random
regression-best linear unbiased prediction model (RR-BLUP) to estimate prediction
accuracies in shoot weight (SW), fresh root weight (FRW), dry matter content
(DMC), and starch yield (SY) in cassava. They showed that using only informative
markers associated with a trait results in higher prediction accuracies for the
respective traits. Prediction accuracies ranging from 0.67 to 0.83 were reported
for SW, FRW, DMC, and SY. Given that prediction accuracies are always less than
1, phenotypic selection (PS) is always more efficient at selection that GS. However,
GS becomes advantageous when considering the shortening of generation cycle
involved in PS. de Oliveira et al. (2012) estimated genetic gains of GS versus PS for
the above traits and concluded that reducing the generation cycle by half with GS
would increase genetic gains by 39.4%, 56.9%, and 73.96% for DMC, FRY, and
SW, respectively. In sugarcane, another complex, polyploid, clonally propagated
crop, Gouy et al. (2013) tested four statistical models, Bayesian LASSO, ridge
regression, reproducing kernel Hilbert space, and partial least square regression,
showing correlations ranging from 0.11 to 0.62 between phenotypes and genotypes
during cross validation, depending on the trait. Equal accuracy was seen for all
models within a trait but with marked differences between traits. They concluded
that their marker density (1499 dominant markers) may not have been large enough
to cover the large sugarcane genome and capture the whole haplotype diversity, and
suggested using multi-allelic markers to improve prediction. In oil palm (Elaeis
guineensis), prediction accuracy ranged from 0.41 to 0.94, depending on the trait
and the relationship of the population to the training set, but was not affected by the
statistical method used (Cros et al. 2015). This was attributed to the small size of the
training population and markers used, as well as to the complex genetic architecture
for different traits. In potato, Habyarimana et al. (2014) used three GS models,
Bayesian LASSO, genomic best linear unbiased prediction (G-BLUP), and
reproducing kernel Hilbert space (RKHS), to evaluate prediction accuracy for
yield, yield components, and quality traits. They reported prediction accuracy of
r > 0.60 for several traits including carotenoids, tuber dry matter, and total yield.
Meanwhile, in apple, prediction accuracy for ten traits (median of 0.19 and a
maximum of 0.5) was strongly affected by distribution of traits and their heritability
(Muranty et al. 2015). Traits with high heritability and normal phenotypic distri-
bution showed response to selection. Furthermore, Kumar et al. (2012) used an 8 K
Illumina Infinium chip to demonstrate the potential of GS for fruit quality traits at
seedling stage, reducing the generation interval for the apple fruit trees with long
juvenile phases. They compared RR-BLUP and Bayesian LASSO methods to show
prediction accuracies ranging from 0.7 to 0.9 according to the trait analyzed but
with little difference between the prediction models. They concluded that GS could
accelerate the breeding process for fruit quality by making selections prior to the
lengthy fruit quality phenotyping. Resende et al. (2012) showed prediction accura-
cies differing for different methods depending on genetic architecture of the traits in
loblolly pine (Pinus taeda). Bayesian methods outperformed RR-BLUP for oligo-
genic traits because RR-BLUP assumes equal contribution of all markers and
can overparameterize by fitting a large number of markers to a trait that is con-
trolled by a few major genes. In other clonally propagated crops like horticultural
and forest trees, sweet potato, yam, and banana, efforts are still underway to
develop and implement GS models and to estimate prediction accuracy for traits
of importance.
8.6 Outlook for Implementing Genomic Selection in Clonal

Crop Breeding Programs
Genomic selection has great potential to expedite the breeding process in clonally
propagated crops by shortening their long breeding cycle. However, the examples
above show relatively moderate prediction accuracies for GS models in different
crops, indicating room for improvement and refinement. Therefore, before the full
potential for GS can be exploited for clonal crops, the challenges associated with
population structure, architecture of traits, polyploidy, rapid LD decay, and hetero-
zygosity have to be addressed in the GS models (Dufresne et al. 2014).
Distinguishing between paralogous copies and the presence of high copy numbers
of repetitive elements is difficult and poses a challenge for full genome annotations
in polyploid crops (Leitch and Leitch 2008). The majority of conventional
SNP-genotyping platforms and analytical tools have been developed for diploid
crops and are often not suited to clonal crops. High heterozygosity and multiple
alleles per locus in many of the clonally propagated crops is a challenge in
developing pipelines for generating the high-density markers necessary for
GS. Further analytical solutions and pipelines are therefore required to allow these
platforms to incorporate partial heterozygosity and allele dosage determination.
Voorrips et al. (2011) developed models that estimate partial heterozygosity in
tetraploid potato. Their methods efficiently assign bi-allelic marker scores in
tetraploid species. Such a method could also be applied to higher ploidies if the
data are of high quality, as more closely spaced peaks are expected at higher
ploidies, which makes efficient assignment to classes a challenge if the data have
too much noise. Serang et al. (2012) developed an algorithm that can estimate
SNP-allele frequencies in individuals with multiple ploidy. They tested the methods
on potato and sugarcane data and found that the methods identified the correct
ploidies for all potato genotypes, whereas a few differences were observed in
sugarcane, in agreement with the unknown ploidy levels of sugarcane genotypes.
These studies are in the right direction but should be validated further in other
polyploid species to allow genetic study of clonally propagated crops to benefit
from next generation sequencing techniques. Once genotype calling and phasing
can be done properly to allow development of high-density markers, efforts should
be put into developing QTL, association mapping, and GS models that account for
all quantitative-genetic (additive, dominant, and epistatic effects) parameters.
Then, proper analysis of such parameters should be performed in available breeding
populations to allow advancement from proof-of-concept status to applied breeding
status.
References
Azevedo CF, de Resende MDV, e Silva FF, Viana JMS, Valente MSF, Resende MFR, Mu~ noz P
(2015) Ridge, Lasso and Bayesian additive-dominance genomic models. BMC Genet 16(1):1
Bisognin DA (2011) Breeding vegetatively propagated horticultural crops. Crop Breed Appl
Biotechnol S1:35–43, Brazilian Society of Plant Breeding
Bradshaw JE (1994) Quantitative genetics theory for tetrasomic inheritance. In: Bradshaw JE,
Mackay GR (eds) Potato Genetics. CABI, Cambridge, pp 71–99
Breiman L (2001) Random forests. Mach Learn 45:5–32
Ceballos H, Kawuki RS, Gracen VE, Yencho GC, Hershey CH (2015) Conventional breeding,
marker-assisted selection, genomic selection and inbreeding in clonally propagated crops: a
case study for cassava. Theor Appl Genet 128:1647–1667
Cros D, Denis M, Sánchez L, Cochard B, Flori A, Durand-Gasselin T, Nouy B, Omoré A,
Pomiès V, Riou V, Suryana E, Bouvet J-M (2015) Genomic selection prediction accuracy in
a perennial crop: case study of oil palm (Elaeis guineensis Jacq.) Theor Appl Genet
128:397–410
Crossa J, Pérez P, Hickey J, Burgue~no J, Ornella L, Cerón-Rojas J, Zhang X, Dreisigacker S,
Babu R, Li Y, Bonnett D (2014) Genomic prediction in CIMMYT maize and wheat breeding
programs. Heredity 112(1):48–60
Denis M, Bouvet JM (2013). Efficiency of genomic selection with models including dominance
effect in the context of Eucalyptus breeding. Tree Genetics & Genomes, 9(1):37–51
D’hoop BB, Paulo MJ, Kowitwanich K, Sengers M, Visser RGF, van Eck HJ, van Eeuwijk FA
(2010) Population structure and linkage disequilibrium unravelled in tetraploid potato. Theor
Appl Genet 121:1151–1170
de Oliveira EJ, Resende MDV, Santos VDS, Ferreira CF, Oliveira GAF, da Silva MS, de Oliveira
LA, Carlos Ivan Aguilar-Vildoso CI (2012) Genome-wide selection in cassava. Euphytica
187:263–276
Dufresne F, Stift M, Vergilino R, Marble BK (2014) Recent progress and challenges in population
genetics of polyploid organisms: an overview of current state-of-the-art molecular and statis-
tical tools. Mol Ecol 23(1):40–69
Flint-Garcia SA, Thornsberry JM, Buckler ES (2003) Structure of linkage disequilibrium in plants.
Annu Rev Plant Biol 54:357–374
Gianola D, Van Kaam JB (2008) Reproducing Kernel Hilbert spaces regression methods for
genomic assisted prediction of quantitative traits. Genetics 178:2289–2303
Gouy M, Rousselle Y, Bastianelli D, Lecomte P, Bonnal L, Roques D, Efile JC, Rocher S,
Daugrois J, Toubi L, Nabeneza S, Hervouet C, Telismart H, Denis M, Thong Chane A,
Glaszmann JC, Hoarau JY, Nibouche S, Costet L (2013) Experimental assessment of the
accuracy of genomic selection in sugarcane. Theor Appl Genet 126:2575–2586
Grattapaglia D, Resende MDV (2011) Genomic selection in forest tree breeding. Tree Genet
Genomes 7:241–255
Griffin PC, Robin C, Hoffmann AA (2011) A next-generation sequencing method for overcoming
the multiple gene copy problem in polyploid phylogenetics, applied to Poa grasses. BMC
biology, 9(1):19
Grüneberg W, Mwanga R, Andrade M, Espinoza J (2009) Breeding clonally propagated crops.
In FAO, selection methods: chapter 13, part 5 ftp://ftp.fao.org/docrep/fao/012/i1070e/
i1070e04.pdf
genome-assisted breeding values. Genetics 177(4):2389–2397
Habier D, Fernando RL, Kizilkaya K, Garrick DJ (2011) Extension of the Bayesian alphabet for
genomic selection. BMC Bioinform 12:186
Habyarimana E, Parisi B, Onofri C, Govoni F, Mandolino G Efficiency of genomic selection for
yield, merceological and nutritional quality traits in hundred thirty-nine cultivates potato
genotype. EAPR 2014 Brussels - 19th triennial conference, 6 to 11 July 2014, conference
paper, doi:10.13140/2.1.1966.9286
Jannink J-L, Lorenz AJ, Iwata H (2010) Genomic selection in plant breeding: from theory to
practice. Brief Funct Genomics 9(2):166–177
Jannoo N, Grivet L, Dookun A, D’Hont A, Glaszmann JC (1999) Linkage desequilibrium among
modern sugarcane cultivars. Theor Appl Genet 99:1053–1060
Jansky SH, Charkowski AO, Douches DS, Gusmini G, Richael C, Bethke PC, Spooner DM, Novy
RG, De Jong H, De Jong WS, Bamberg JB (2016) Reinventing potato as a diploid inbred line-
based crop. Crop Sci 56:1–11
Jones A (1965) Cytological observations and fertility measurements of sweetpotato [Ipomoea
batatas (L.) Lam]. Proc Am Soc Horticult Sci 86:527–537
Kumar S, Chagné D, Bink MCAM, Volz RK, Whitworth C, Carlisle C (2012) Genomic selection
for fruit quality traits in apple. PloS One 7(5):e36674
Kumar S, Garrick DJ, Bink MCAM, Whitworth C, Chagné D, Richard K, Volz RK (2013) Novel
genomic approaches unravel genetic architecture of complex traits in apple. BMC Genomics
2013(14):393
Leitch A, Leitch I (2008) Genomic plasticity and the diversity of polyploid plants. Science
320:481–483
Ly D, Hamblin M, Rabbi I, Melaku G, Bakare M, Gauch HG Jr, Okechukwu R, Dixon AGO,
Kulakow P, Jannink J-L (2013) Relatedness and genotype environment interaction affect
prediction accuracies in genomic selection: a study in cassava. Crop Sci 53:1312–1325
Meuwissen T (2007) Genomic selection: marker-assisted selection on a genome-wide scale. J
Anim Breed Genet 124(6):321–322
Miller AJ, Gross BL (2011) From forest to field: perennial fruit crop domestication. Am J Bot 98
(9):1389–1414
Mu~noz PR, Resende MFR, Gezan SA, Resende MDV, de los Campos G, Kirst M, Huber D, Peter
GF (2014) Unraveling additive from non-additive effects using genomic relationship matrices.
Genetics 198:1759–1768
Muranty H, Troggio M, Sadok IB, Rifai MA, Auwerkerken A, Banchi E, Velasco R, Stevanato P,
van de Weg WE, Di Guardo M, Kumar S, Laurens F, Bink MCAM (2015) Accuracy and
responses of genomic selection on key traits in apple breeding. Horticult Res 2:15060. doi:10.
1038/hortres.2015.60
Myles S (2013) Improving fruit and wine: what does genomics have to offer? Trends Genet 29
(4):190–196
Nakaya A, Isobe SN (2012) Will genomic selection be a practical method for plant breeding? Ann
Bot 110:1303–1316
Park T, Casella G (2008) The Bayesian lasso. J Am Stat Assoc 103:681–686
Poland J, Rutkoski J (2016) Advances and challenges in genomic selection for disease resistance.
Annu Rev Phytopathol 54:79–98
Raboin L-M, Pauquet J, Butterfield M, D’Hont A, Glaszmann J-C (2008) Analysis of genome-
wide linkage disequilibrium in the highly polyploid sugarcane. Theor Appl Genet 116:701–714
Resende MFR, Mu~noz P Jr, Resende MDV, Garrick DJ, Fernando RL, Davis JM, Jokela EJ,
Martin TA, Peter GF, Kirst M (2012) Accuracy of genomic selection methods in a standard
data set of loblolly pine (Pinus taeda L.) Genetics 190:1503–1510
Sardos J, Rouard M, Hueber Y, Cenci A, Hyma KE, van den Houwe I et al (2016) A genome-wide
association study on the seedless phenotype in banana (Musa spp.) reveals the potential of a
selected panel to detect candidate genes in a vegetatively propagated crop. PLoS One 11(5):
e0154448
Serang O, Mollinari M, Garcia AAF (2012) Efficient exact maximum a posteriori computation for
Bayesian SNP genotyping in polyploids. PLoS One 7:e30906
Simko I, Haynes KG, Jones RW (2006) Assessment of linkage disequilibrium in potato genome
with single nucleotide polymorphism markers. Genetics 173:2237–2245
Simmonds NW (1962) The Evolution of the Bananas. Tropical Sciences Series. Longman,
London, pp. 170
Simmonds NW (1979) Principles of crop improvement. Longman, New York
Slater AT, Cogan NOI, Forster JW, Hayes BJ, Daetwyler HD (2016) Improving genetic gain with
genomic selection in autotetraploid potato. Plant Genome 9. doi:10.3835/plantgenome2016.02.
0021
Spindel J, Begum H, Akdemir D, Virk P, Collard B, Redo~ na E et al (2015) Genomic selection and
association mapping in rice (Oryza sativa): effect of trait genetic architecture, training popu-
lation composition, marker number and statistical model on accuracy of rice genomic selection
in elite, tropical rice breeding lines. PLoS Genet 11(2):e1004982
Spooner DM, Gavrilenko T, Jansky SH, Ovchinnikova A, Krylova E, Knapp S, Simon R (2010)
Ecogeography of ploidy variation in cultivated potato (Solanum sect. Petota). Am J Bot 97
(12):2049–2060
Stich B, Urbany C, Hoffmann P, Gebhardt C (2013) Population structure and linkage disequilib-
rium in diploid and tetraploid potato revealed by genome-wide high density genotyping using
the SolCAP SNP array. Plant Breed 132:718–724
van Nocker S, Gardiner SE (2014) Breeding better cultivars, faster: applications of new technol-
ogies for the rapid deployment of superior horticultural tree crops. Horticult Res 1:14022.
doi:10.1038/hortres.2014.22
Voorrips R, Gort G, Vosman B (2011) Genotype calling in tetraploid species from bi allelic marker
data using mixture models. BMC Bioinformatics 12:172
Weng Z, Wolc A, Shen X, Fernando RL, Dekkers JC, Arango J et al (2016a) Effects of number of
training generations on genomic prediction for various traits in a layer chicken population.
Genet Sel Evol 48(1):1
Weng Z, Wolc A, Shen X, Fernando RL, Dekkers JCM, Arango J, Settar P, Fulton JE, O’Sullivan
NP, Garrick DJ (2016b) Effects of number of training generations on genomic prediction for
various traits in a layer chicken population. Genet Sel Evol 48:22
Windhausen VS, Atlin GN, Hickey JM, Crossa J, Jannink J-L, Sorrells ME, Raman B, Cairns JE,
Tarekegne A, Semagn K, Beyene Y, Grudloyma P, Technow F, Riedelsheimer C, Melchinger
AE (2012) Effectiveness of genomic prediction of maize hybrid performance in different
breeding populations and environments. G3 Genes Genomes Genet 2:1427–1436
Wolfe MD, Rabbi IY, Egesi C, Hamblin M, Kawuki R, Kulakow P, Lozano R, Del Carpio DP,
Ramu P, Jannink J-L (2016) Genome-wide association and prediction reveals genetic archi-
tecture of cassava mosaic disease resistance and prospects for rapid genetic improvement.
Plant Genome 9. doi:10.3835/plantgenome2015.11.0118
Yan J, Shah T, Warburton ML, Buckler ES, McMullen MD et al (2009) Genetic characterization
and linkage disequilibrium estimation of a global maize collection using SNP markers. PLoS
One 4(12):e8451. doi:10.1371/journal.pone.0008451
Chapter 9
Status and Perspectives of Genomic Selection
in Forest Tree Breeding
Dario Grattapaglia
9.1 Introduction: From Trait Dissection to Genomic

Prediction in Forest Trees
Advanced tree breeding involves a large number of steps around the basic concept
of increasing the frequency of favorable alleles for a number of traits simulta-
neously in the target population. Recurrent cycles of selection, mating, and testing
are used to develop genetically improved seeds or elite clonal stocks by maximizing
genetic gain per unit time at the lowest possible cost (Namkoong et al. 1988; White
et al. 2007). Trees have long life cycles and become reproductively active only after
several years. The progress and success of tree breeding programs are therefore
strongly dependent on the time needed to complete a breeding generation. This may
last several years to decades depending on the biology of the species, the age at
which phenotypes can be accurately measured, and the deployment plan of
improved material, whether seeds or clones. Additionally, the uncertainties associ-
ated with conducting decade-long breeding programs can be high. Breeding invest-
ments are made several years before the eventual utilization of genetically
improved material, making it susceptible to changes in the economic objectives
of the forest products, market demands, and management policies.
The time challenges faced by tree breeders have historically led to substantial
efforts to understand juvenile-mature correlations for late-expressing traits
(Namkoong et al. 1988), devise ways to accelerate recombination by artificial
flower induction (Greenwood et al. 1991; Hasan and Reid 1995), and practice
early selection on juvenile traits (Williams 1988). In the early 1990s, when DNA
marker technologies became more accessible, marker-assisted selection (MAS) was
D. Grattapaglia (*)
EMBRAPA Genetic Resources and Biotechnology – EPqB, 70770-910 Brasilia, DF, Brazil
Universidade Católica de Brası́lia- SGAN, 916 modulo B, Brası́lia, DF 70790-160, Brazil
e-mail: dario.grattapaglia@embrapa.br

DOI 10.1007/978-3-319-63170-7_9
200 D. Grattapaglia
immediately seen as a powerful tool to overcome some of the challenges. MAS

could shorten breeding cycles by early selection for late-expressing traits such as
wood properties. Furthermore, MAS could also be applied to increase selection
intensity, reduce the effort of field-testing, and possibly improve selection precision
for low-heritability traits such as volume growth (Grattapaglia et al. 1992; Neale
and Williams 1991; Williams and Neale 1992). Nevertheless, the potential of MAS
for forest trees was immediately questioned based on the state of linkage equilib-
rium that the typically large, random mating population with recent domestication
history would be found and the concerns regarding stability of QTLs across the
highly variable genetic backgrounds of breeding populations and environments
(Strauss et al. 1992).
Despite those early, well-grounded arguments, linkage mapping and QTL detec-
tion experiments in species of pines, spruces, poplars, and eucalypts were carried
out based on the implicit assumption that it would be possible to map and estimate
the effects of all the relevant genes for traits such as growth and wood quality
during the life of the tree, in every population and environment. A considerable
number of studies describing QTLs and gene-trait associations in forest trees were
reported (reviewed in (Grattapaglia et al. 2009, 2012; Harfouche et al. 2012; Neale
and Kremer 2011). Mirroring what was the canonical approach to QTL mapping in
the major crops and model systems, QTL mapping in forest trees was carried out
using single biparental populations of relatively limited size. The difference was
that two-generation pedigrees were used because any cross between heterozygous
parents would provide a segregating F1 population under a pseudo-testcross
(Grattapaglia and Sederoff 1994). Several “major effect” QTLs were mapped in
early studies. However, later multifamily experiments conducted with larger
populations revealed many more QTLs with smaller effects and largely inconsistent
across backgrounds and environments (Dillen et al. 2008; Freeman et al. 2013;
Gion et al. 2011; Novaes et al. 2009; Rae et al. 2008; Thumma et al. 2010). With the
exception of a few QTLs of moderately large effect mapped for disease resistance
(Junghans et al. 2003; Stirling et al. 2001; Wilcox et al. 1996), and candidate-gene
associations for phenological traits (Ingvarsson et al. 2008), results generally
showed that QTL and association mapping do not explain sufficient genetic vari-
ation to lead to any effective implementation of MAS for complex traits in forest
trees. On hindsight, it is perplexing to consider how far removed from the reality of
forest tree breeding were those biparental mapping populations and the QTL
mapping data derived from them.
The ineffectiveness in dissecting complex traits, i.e., determining the position,
variation, and magnitudes of allelic effects at QTLs underpinning quantitative traits
and the consequent failure to implement MAS, has not been exclusive to forest
trees. With the exception of a few simple qualitative or monogenic traits in crops
(Bernardo 2008), major genes in fruit trees (Arus et al. 2012), and recessive genetic
defects in domestic animals (Charlier et al. 2008), this has been the general
conclusion for the vast majority of species undergoing selective breeding. This
fact has caused a major paradigm shift in animal and plant molecular breeding in
the last 10 years. The field has now moved from trying to a priori discover, validate,
9 Status and Perspectives of Genomic Selection in Forest Tree Breeding 201
and use marker-trait associations to dealing with the aggregate of the whole-
genome effect, much like quantitative genetics always did. This revolution was
only possible following the development of large numbers of markers, mostly
SNPs, together with cost-effective platforms to query them genome-wide and
new statistical methods to deal with large datasets accounting for the large numbers
of markers ( p) and relatively small number of individuals (n) problem. The
approach termed “genomic selection” (GS) estimates all marker effects simulta-
neously, retaining all of them as predictors of performance and precluding the prior
search for significant marker-trait associations but focusing exclusively on predic-
tion efficiency (Goddard and Hayes 2009; Meuwissen et al. 2001).
Genomic selection has become a theme of considerable interest in the tree
genetics and breeding community worldwide in the last few years since the first
perspectives based on simulations (Grattapaglia and Resende 2011; Iwata et al.
2011) and experimental results (Resende et al. 2012a, b) were reported. In this
chapter, an update is provided of an earlier review on this topic (Grattapaglia 2014).
However, a more comprehensive discussion of the main factors (theoretical and
practical) relevant to GS in tree breeding that has emerged from experimental
studies in the last few years is provided. This discussion is preceded by a concise
explanation of the basic insights of GS and its perspectives and challenges in tree
breeding. An updated compilation of all published experimental GS studies in
forest trees follows, highlighting their main contributions to our current under-
standing of this new approach for tree breeding. The conclusion finally summarizes
the main lessons learned so far in an attempt to provide a nine-point tentative
roadmap for implementing GS in a tree breeding program.
9.2 Genomic Selection: Reviewing the Basic Principles
Nejati-Javaremi et al. (1997) were probably the first ones to show that “total allelic”
relationship estimated from marker data would be a powerful alternative to the
pedigree-derived additive genetic relationship to derive best linear unbiased pre-
diction estimates (EBV) of breeding values using mixed model equations. Haley
and Visscher (1998) proposed the idea of “total genomic selection,” that is, that by
genotyping at the genome-wide scale with sufficient marker density and low cost, it
would be possible to assure that markers will be in complete association with any
trait locus and, therefore, capture the most genomic effects underlying complex
traits. However, it was the groundbreaking paper by Meuwissen et al. (2001) that
anticipated that selection on genetic values predicted from markers could consid-
erably increase the rate of genetic gain in animal and plant breeding programs. They
also outlined the statistical approach and potential caveats to estimate the genetic
value of unphenotyped individuals based exclusively on phenotype and genotype
data of a reference ancestral population using “genomic selection” (GS), a term
surprisingly not used in the main text but only in the running title of that paper.
202 D. Grattapaglia
Both MAS and GS start by establishing associations between discrete marker

genotypes and continuously distributed phenotypes in relevant populations. How-
ever, they are fundamentally different following this initial step. MAS typically
targets the discovery of marker-trait associations in one or a few biparental
populations or association mapping panels using rigorous significance tests, with
the later goal of using such marker-trait associations for selection. GS instead uses a
dense genome-wide panel of markers whose effects on the phenotype are estimated
simultaneously in a large and representative “training” population of individuals
without applying rigorous significance tests. All or a subset of markers are retained
as forecasters of phenotypes in prediction models to be later applied to “selection
candidates” for which only genotypes are collected. Thus, in GS a marker effect
does not need to exceed a stringent significance threshold to be used in the
subsequent breeding phase, and the effects of the marker alleles are estimated in
a much larger and more representative population rather than within one or a few
mapping families. The “training” population involves at least several hundreds to a
few thousand individuals representative of the target breeding population, which
are genotyped for the marker panel and phenotyped for all traits of interest. The
prediction models developed for each trait are cross-validated in a “validation”
population, a randomly sampled subset of individuals of the same reference pop-
ulation that did not participate in the estimation of marker effects. Once a prediction
model is shown to provide satisfactory accuracy, i.e., correlation between the
observed and predicted breeding values following cross-validation, it can be used
in the breeding phase to calculate the genomic estimated breeding values (GEBV)
or total genomic estimated genotypic values (GEGV) (when nonadditive effects are
also included in the model) of the selection candidates. Put simply, a GEBV is
calculated by multiplying the genotypes at all markers by their effect estimated by,
for example, random regression best linear unbiased prediction (RR-BLUP) or any
other statistical method that adequately avoids model over-fitting by marker-
specific shrinkage of regression coefficients (Crossa et al. 2010; Lorenz et al.
2011). There is also a second approach to use genotypic data in GS. Marker
genotypes are used to estimate a genomic relationship matrix between individuals
with genotypes and phenotypes of the training population and the yet-to-be
phenotyped selection candidates for which only genotypes are available. This
genomic relationship matrix can then be used to estimate a variance/covariance
matrix between the genetic values in a mixed model generally called G-BLUP that
stands for genomic BLUP. It has been shown that RR-BLUP and G-BLUP are
statistically equivalent under theoretical conditions that are generally met in prac-
tice (Habier et al. 2007).
GS exploits both the linkage disequilibrium (LD) between the dense marker data
and all QTL effects associated and the genetic relationship between the training
population and the prospective selection candidates. By avoiding prior marker
selection and estimating marker effects in a large and representative population,
GS potentially captures all genetic variance for the trait explained by the large
numbers of small effects that QTL or association genetics-based MAS does not
capture. Genomic selection also known as genome-wide selection (GWS) has now
become the paradigm for marker-assisted selection (MAS) of complex traits in

plants and animals. GS is the standard molecular breeding technology in dairy cattle
and increasingly been adopted in other animal species such as swine and broiler
(Van Eenennaam et al. 2014). The most extraordinary example continues to be in
dairy cattle where progeny testing of young bulls has been replaced by GS, resulting
in rapid improvements across multiple traits. By 2011 over 40% of the market share
of tested bulls across several countries was composed of bulls without milking
daughters, exclusively selected based on GS (Pryce and Daetwyler 2012).
While GS was rapidly being adopted in animal breeding, it also became a topic
of interest in plants, starting with the influential papers by Bernardo (2008) and
Bernardo and Yu (2007), soon followed by others that discussed the potential of
genomic prediction in crops (Heffner et al. 2009; Jannink et al. 2009) and forest tree
improvement (Grattapaglia et al. 2009). An exponential growth of published studies
about GS in all major cultivated plant species has taken place in the last 5 years.
Following the early enthusiasm and prospects fueled by several simulation-based
studies validated by experimental results, we have now moved to a phase where
several detailed and careful considerations are necessary (Heslot et al. 2015; Jonas
and de Koning 2015). These include the strategic breeding and tactical logistics and
resource allocation aspects of implementing GS, the issues related to the optimal
planning of training populations and phenotyping efforts associated to them, the
marker platforms to be used, and a thorough cost-benefit analysis of the entire
process.
9.3 Perspectives of Genomic Selection in Tree Breeding
The objective of selective breeding is to accelerate the rate of genetic improvement

or selection response per unit time. As noted above, the time factor is extremely
relevant to tree breeding due to the long generation times typically necessary to
complete a full breeding cycle. To go back to the classic breeder’s equation is
therefore useful to understand how GS can have a tremendous impact on the rate of
genetic gain. In the equation (ΔG ¼ irσ A/L), i is the selection intensity (the
proportion of trees that are selected to become parents of the next generation);
r is the accuracy of selection, i.e., the correlation between the estimated breeding
value (EBV) and the true breeding value; σ A is the additive genetic standard
deviation of the trait of interest, i.e., the genetic variation available in the population
for selection; and L is the generation interval or time needed to achieve the genetic
gain. GS can directly increase the rate of genetic gain of a tree breeding program by
increasing the selection intensity (i), because many more young seedlings can be
genotyped and their phenotypes predicted by GS than the number of seedlings
typically planted and managed in field trials. This is particularly relevant for traits
expensive to measure or expressed late in the life of the tree. However, the largest
impact of GS on the rate of genetic gain will result from radically reducing the
generation interval (L). Phenotypes of the selection candidates can be predicted at
204 D. Grattapaglia
ultra-early ages, for example, when the seedlings are a few weeks old, still in the
nursery, instead of waiting half of the breeding cycle (usually 4–20 years or more,
depending on the species) before having access to their phenotypes, especially
those expressed late in the life of the tree. The accuracy of selection (r) evidently is
also a main driver of the genetic gain. In standard breeding this accuracy is provided
by the square root of the heritability, i.e., the proportion of the phenotypic variance
explained by genetic components. In GS the accuracy of selection is estimated by
the correlation between the genomic estimated breeding value (GEBV) and the true
breeding value. When the breeding objective is to select individuals to be deployed
as clones, this correlation needs to involve not only the additive effects but also the
nonadditive component, such that r measures the correlation between the genomic
estimated genotypic value (GEGV) and the true genotypic value. Finally, without
genetic variation (σ A) for the target trait, no progress will happen.
The implementation of a genomic selection program for tree breeding encom-
passes essentially two stages (Fig. 9.1). The first one involves the definition of a
“training population” of individuals that are genotyped and phenotyped to develop
and cross-validate predictive models to be later used in the second stage, where GS
is actually put in practice. A training population is usually sampled from an existing
progeny trial derived from inter-mating a group of elite parents that were
established as the population to undergo breeding for the subsequent generations.
Usually this group of elite parents will have an effective population size (Ne)
between 30 and 100 and a census number (N ) that will be in that same range or
slightly larger, taking into account any cryptic relatedness that exists between the
individuals. The training population will have at least 1,000–2,000 individuals.
However, the more individuals are genotyped and phenotyped, the better will the
marker effects be estimated and more robust will become the predictive model.
In the second stage, GS will be effectively employed on the selection candidates,
typically an array of full of half-sib families derived from intercrossing either the
original elite parents or elite individuals selected in the progeny trial used as
training population. These selection candidates are genotyped and have their
breeding values (GEBV) and/or genotypic values (GEGV) estimated using the
predictive model developed earlier. Top ranked seedlings for GEBV are subject
to early flower induction and inter-mated to create the next generation of breeding.
Top ranked seedlings for GEGV are clonally propagated and tested in verification
clonal trials where elite clones are eventually selected for operational plantation.
Additionally, a random subset of the already genotyped selection candidates could
be planted in experimental design and phenotyped at the target age to provide
genotype and trait data for GS model updating as generations of GS advance,
mitigating the erosion of marker-QTL LD and decay of relationships, and
maintaining accuracy of GS predictions over generations.
Simulation-based and experimental reports outlined the promising prospects of
GS to increase the efficiency of tree breeding programs (see below). In eucalypts
and poplars, GS not only could eliminate the progeny trial but would also reduce the
time and costs involved in the clonal testing phase by reducing the number of
selected trees that are evaluated as clones in a preliminary, typically large-scale,
9 Status and Perspectives of Genomic Selection in Forest Tree Breeding
Fig. 9.1 The two stages of the development and implementation of a genomic selection program in tree breeding (Modified from Grattapaglia (2014))
205
206 D. Grattapaglia
clonal trial. In conifers, as pointed out by Resende et al. (2012b), GS combined to

somatic embryogenesis (SE) could considerably boost the efficiency of current
clonal propagation protocols by allowing preselection of zygotic embryos based
on their GEBV and their immediate expansion into elite SE lines for the establish-
ment of clonal trials or directly into commercial plantations. Besides the time gain,
a less mentioned advantage of GS is related to the possibility of efficiently carrying
out selection for several traits simultaneously in large numbers of individuals. It is
virtually impossible for any breeding program to complete a rigorous assessment of
all traits of interest in all trees in a progeny trial. Traits usually include wood
volume, stem taper and straightness, physical and chemical wood properties,
sprouting and rooting abilities, nutritional efficiency, and tolerance to pests, dis-
eases, drought, and frost. In tropical eucalypts, for example, even in clonal trials,
this is typically accomplished only in the very final stages of selection and for a very
limited number of clones (20–50) that had been preselected for volume growth and
wood density (Rezende et al. 2014). In traditional breeding, a sequential approach is
typically used that combines different forms of selection indices and independent
culling levels for estimating the ultimate value of candidates. In GS, because
breeding values are predicted for each trait separately (i.e., a separate GEBV for
each measured trait), selection indices can be used to combine data from all the
traits under analysis into a single value for each candidate. The validity of this
multiple-step approach rests on a property that the BLUP of any linear combination
of traits is equal to that linear combination of the BLUP predicted values of the
individual traits (White et al. 2007). Therefore, the net effect of GS would be a
remarkable increase in selection intensity at the seeding stage for all traits simul-
taneously, considerably improving the overall efficiency of the breeding program.
Additionally, applying GS to multiple traits could significantly increase the predic-
tion accuracy for a low-heritability trait or for traits with a limited number of
phenotypic records when a correlated high-heritability trait is available (Jia and
Jannink 2012).
9.4 Genomic Selection: Experimental Results in Forest

Trees
A compilation of all experimental GS studies in forest trees published to date is

provided in a format that allows a quick perusal of their key attributes and
performance of predictive abilities for different traits (Table 9.1). Reports of GS
in forest trees have been unique in that they used considerably larger training
population sizes and numbers of markers when compared to GS studies in crop
plants. Experiments have typically mirrored the structure of true breeding
populations and adopted designs that accounted at satisfaction with the theoretical
expectations of higher diversity and the necessary relationship between training and
validation sets. Another distinctive aspect has been the attempt to evaluate the
Table 9.1 Summary of the main features and results of the published experimental studies of genomic selection in forest trees
Population Predictive
Species Population structure size # and type of markers used Trait (h2)a abilityb Reference
Eucalypts Progeny trial of 43 full-sib 738 3,129 DArT fixed array DBH (0.53) 0.54 Resende et al.
(Eucalyptus families from 11 elite hybrid HG (0.42) 0.51 (2012a)
grandis x parents WSG (0.59) 0.60
E. urophylla
PY (0.38) 0.54
hybrids)
Eucalypts Progeny trial of 232 full-sib 920 3,564 DArT fixed array DBH (0.56) 0.55
(E. grandis, families from 51 elite parents HG (0.48) 0.46
E. urophylla, E. WSG (0.42) 0.42
globulus, and
PY (0.47) 0.38
their F1 hybrids)
Loblolly pine Progeny trial of 61 full-sib 800 4,852 SNP DBH (0.21–0.32c) 0.31–0.37c Resende et al.
Pinus taeda families from 32 parents repli- Infinium chip HG (0.13–0.26a) 0.26–0.34c (2012b)
cated in four environments;
trees were clonally replicated
Loblolly pine Progeny trial of 61 full-sib 951 4,853 SNP Infinium chip GR (0.22–0.35) 0.38–0.49 Resende et al.
Pinus taeda families from 32 elite parents; DVP (0.07–0.45) 0.24–0.51 (2012c)
trees were clonally replicated FRR (0.12–0.21) 0.23–0.34
WS (0.37) 0.39–0.43
LC (0.11) 0.17
LTW (0.17) 0.23–0.24
WSG (0.09) 0.20–0.22
SC (0.14) 0.25–0.26
Loblolly pine 13 full-sib families related by 149 3,406 SNP Infinium chip LC (0.76)d 0.66–0.76 Zapata-
Pinus taeda common parents CC (0.69) 0.61–0.83 Valenzuela et al.
HG (0.95) 0.47–0.52 (2012)
VG (0.94) 0.30–0.56
Loblolly pine 13 full-sib families related by 165 3,461 SNP Infinium chip HG 0.37–0.74e Zapata-
Pinus taeda common parents Valenzuela et al.
(2013)
207
Loblolly pine Progeny trial of 61 full-sib 951 4,853 SNP Infinium chip HG (0.32) 0.66–0.86f Munoz et al.
Pinus taeda families from 32 elite parents (2014)
(continued)
208
White spruce 214 open-pollinated families 1,694 6,385 SNPs Infinium chip CP (0.24) 0.39; 0.20; (Beaulieu et al.
Picea glauca from 43 natural populations 0.09g 2014a)
FC (0.33) 0.39; 0.18;
0.13
CW (0.57) 0.33; 0.17;
0.0
WSG (0.39) 0.37; 0.13;
0.07
MA (0.38) 0.38; 0.16;
0.0
WS (0.31) 0.38; 0.18;
0.03
RW (0.04) 0.33; 0.13;
0.0
SFS (0.48) 0.39; 0.16;
0.12
CRD (0.44) 0.44; 0.28;
0.13
CTD (0.26) 0.35; 0.15;
0.11
CWT (0.39) 0.41; 0.13;
0.13
HG (0.25) 0.36; 0.18;
0.09
D. Grattapaglia
White spruce Two unrelated breeding groups 1,748 6,932 SNP Infinium chip WSG (0.33–0.43)h 0.79–0.06i Beaulieu et al.
Picea glauca with 851 and 897 individuals in 0.75–0.66j (2014b)
27 and 32 full-sib families, 0.59–0.10k
respectively, in two 0.63–0.53l
environments MA (0.30–0.32) 0.77–0.03
0.71–0.61
0.36–0.04
0.52–0.41
HG (0.39–0.57) 0.58–0.0
0.68–0.34
0.33–0.0
0.55–0.22
DBH (0.32–0.39) 0.52–0.0
0.69–0.24
0.29–0.0
0.48–0.15
Eucalypts Progeny trial of 45 full-sib 1,000 29,090 SNPs from the DBH (0.64) 0.44 Lima (2014)
(Eucalyptus families from 46 elite parents EuCHIP60K fixed array HG (0.54) 0.34
grandis x VG (0.63) 0.42
E. urophylla
MAI (0.63) 0.42
hybrids)
CC (0.74) 0.52
HC (0.81) 0.55
SGR (0.93) 0.83
ILC (0.82) 0.64
SLC (0.84) 0.72
LC (0.84) 0.63
WSG (0.76) 0.63
MA (0.26) 0.17
FL (0.80) 0.49
FW (0.22) 0.14
FC (0.49) 0.33
209
(continued)
210
Interior spruce 25 open-pollinated families in 1,126 8,868 to 62,198 SNPs by GbS HG (0.43-0.98)h 0.17–0.63m El-Dien et al.
Picea glauca x three environments with different imputation DBH(0.28-0.55) 0.01–0.77 (2015)
Picea methods to account for 30% or VG (0.29-0.76) 0.01–0.73
engelmannii 60% missing data
WDV(0.31-0.78) 0.17–0.67
WDR(0.42-0.65) 0.14–0.64
WDX(0.48-0.59) 0.23–0.62
MOD(0.31-0.78) 0.16–0.67
Interior spruce 25 open-pollinated families in 769 34,570 to 50,803 SNP by GbS HG (0.43-0.98) 0.04–0.47 Ratcliffe et al.
Picea glauca x two environments depending imputation method (2015)
Picea
engelmannii
Maritime pine 184 founders (G0) and 661 2,500 SNPs Infinium chip DBH (0.17) 0.38–0.47 Isik et al. (2016)
Pinus pinaster 477 progeny individuals HG (0.30) 0.45–0.47
(G1) in 191 maternal half-sib SS (0.25) 0.48–0.55
families
D. Grattapaglia
Maritime pine 46 founders (G0), 62 (G1) and 818 4,332 SNPs Infinium chip DBH (0.17) 0.52–0.74 Bartholome et al.
Pinus pinaster 710 progeny individuals HG (0.30) 0.58–0.69 (2016)
(G2) in 35 maternal half-sib SS (0.25) 0.65–0.82
families but with full pedigree
inferred by SNP data
a
Trait heritability estimated using pedigree data; trait legend: HG height growth, DBH diameter at breast height, WSG wood-specific gravity, PY pulp yield,
LC lignin content, CC cellulose content, GR growth (several traits), DVP developmental traits, FRR fusiform rust resistance, WS wood stiffness, LTW
latewood %, SC sugar content, VG volume growth, CP cell population, FC fiber coarseness, CW crystallite width, MA microfibril angle, RW ring width, SFS
specific fiber surface, CRD cell radial diameter, CTD cell tangential diameter, CWT cell wall thickness, SS stem sweep, WDV wood density by acoustic
velocity, WDR wood density by resistance to drilling, WDX wood density by X-ray densitometry, MOD modulus of elasticity, HC hemicellulose content, SGR
syringyl/guaiacyl lignin ratio, ILC insoluble lignin content, SLC soluble lignin content, FL fiber length, FW fiber width
b
Predictive abilities reported correspond to the correlation between observed and expected breeding values; GRM genomic relationship matrix, PA predictive
ability
c
Heritability or PA variation across the four environments evaluated
d
Heritabilities were reported as clone mean repeatabilities
e
Variation in PA depending on the cross-validation method used
f
PAs of the additive model and additive*dominance model
g
PAs estimated by three different cross-validation schemes with decreasing relatedness between training and validation (see text for further details)
h
Heritabilities across the different environments tested
i
PAs from within-families cross-validation within and between the two breeding groups
j
PAs from within-families cross-validation within and between the two environments
k
PAs from between-families cross-validation within and between the two breeding groups
l
PAs from between-families cross-validation within and between the two environments
m
PAs reported are the lowest estimates by cross-validation across environments and the highest multisite using all individuals for training and validation
211
212 D. Grattapaglia
impact of different issues particularly relevant to tree breeding on the accuracy of

prediction. These included the level of relationship between training and validation
sets, the effect of genotype by environment (G*E), the influence of age-age
correlations, and the performance of different analytical approaches that use vari-
able underlying assumptions of trait architecture. Nevertheless, all studies until
recently were not able to evaluate the actual performance of GS across generations,
i.e., using training data of an ancestral generation to predict and validate phenotypes
of progeny individuals. Cross-validation and estimation of predictive accuracy was
carried out exclusively within the same generation. Recently, however, studies with
Pinus pinaster that had access to three generations (G0, G1, and G2) showed
encouraging results of intergeneration prediction both by mixing parents and
progeny in the same training set (Isik et al. 2016) and later using only G0 and G1
individuals to predict in the G2 generation (Bartholome et al. 2016). Reported
prediction accuracies of experimental studies have been generally very satisfactory,
in line with the expectations from previous simulations (Grattapaglia and Resende
2011; Iwata et al. 2011) and results in crop plants and domestic animals.
Two experimental studies pioneered the field of genomic prediction in forest
trees. A report in Eucalyptus involving two independent genetically unrelated
breeding populations with contrasting effective population sizes assessed in
completely different environments (Grattapaglia et al. 2011b; Resende et al.
2012a) and a second one involving a cloned set of loblolly pine full-sib families
assessed across four different environments and two different ages (Resende et al.
2012b). Predictive abilities between 0.26 and 0.60, with an overall average of 0.44,
were estimated by cross-validation for a range of growth and wood quality traits.
These results approximated well to the accuracies predicted from deterministic
(Grattapaglia and Resende 2011) and stochastic simulations (Iwata et al. 2011) for
similar parameters of trait heritability, effective population size, and genotyping
density. These experimental results suggested that potential gains of 50–200% in
selection efficiency predicted by simulations could be achieved, if adequate pre-
diction abilities would be kept across generations. These studies also showed that
prediction accuracies strongly depend on the existence of genetic relationship
between training and validation sets and are impacted by G*E and age-age corre-
lations such that predictions will be effective when carried out in the same envi-
ronment and same age as where and when the training data was collected. Studies in
white spruce soon followed where the impact of G*E and the key significance of
having genetic relationships between training and validation were thoroughly
evaluated and corroborated (Beaulieu et al. 2014a, b). Recently published reports
in eucalypts (Lima 2014), spruce (El-Dien et al. 2015; Ratcliffe et al. 2015), and
maritime pine (Bartholome et al. 2016; Isik et al. 2016) provided additional
promising results on the ability to predict complex traits in forest trees and
confirmed what had been previously observed as far as the impact of relationship,
G*E, and age. The main results and contributions of all these studies are detailed
below, while discussing the main factors that affect the prospects of GS to forest
tree breeding.
9.5 Factors Affecting the Success of Genomic Selection

in Tree Breeding
The success of GS is dependent on a number of factors including both the funda-

mental aspects predicted from population and quantitative genetics theory and the
more practical and logistics aspects of resource allocation and cost-benefit analysis.
The following discussion tries to cover the main factors in these two realms
reminding that they are in many ways interconnected and interdependent. The
accuracy of a genomic prediction model, i.e., the correlation between the genomic
estimated breeding value (GEBV) and the true breeding value, is undoubtedly the
key factor that will have the major impact on the success of GS. Four fundamental
factors from the theory of population and quantitative genetics are known to affect
the accuracy of genomic prediction: (1) the effective population size (Ne) and
genotyping density that in turn determines the extent of LD between markers and
QTLs; (2) the size and composition of the training population, i.e., the number of
individuals with phenotypes and genotypes from which the marker effects are
estimated; (3) the heritability of the trait in question; and (4) the genetic architecture
of the target trait, i.e., the distribution of QTL effects (number of loci and size
effects) (Hayes et al. 2009a; Lin et al. 2014). An assessment of the impact of each
one individually in the context of tree breeding was reported early on, providing
some broadly useful guidelines for GS regardless of the target species, recombinant
genome size, or breeding cycle length (Grattapaglia and Resende 2011).
When considering the practicalities of tree breeding, some of the most relevant
factors that impact the prospects of GS include (1) the size, composition, and
phenotyping effort devoted to the training population; (2) the genotyping platform
employed and the resulting data quality, cost, turn-around time, and breeder’s
friendliness; (3) the extent of genotype by environment interaction; and (4) the
long-term performance of genomic prediction models, including the need for model
retraining and the potential effect on loss of diversity and increased inbreeding.
With all these issues considered, an attempt is made here to answer a common
question posed by tree breeders: what are the main issues that one should be aware
of when considering the investment in a GS program?
9.5.1 Effective Population Size of the Tree Breeding

Population
The main issue generally considered to determine the accuracy of GS is the extent
of linkage disequilibrium, i.e., the nonrandom association between marker alleles
and QTL alleles. This factor in turn directly depends on the effective population
size (Ne) and genotyping density. The effective population size corresponds to the
number of breeding individuals in an idealized population that would show the
same amount of dispersion of allele frequencies under random genetic drift or the
214 D. Grattapaglia
same amount of inbreeding as the population under consideration (Wright 1931).

As the effective population size gets smaller, the effect of genetic drift gets
stronger, and more LD is generated because it is unlikely that combinations of
marker alleles and QTL alleles get sampled at a frequency that corresponds to the
product of their individual frequencies. The resulting nonrandom association
between alleles at marker loci and QTLs allows marker alleles to predict the allelic
state of nearby QTL and thus to predict phenotypes. At equilibrium, the LD
generated by random drift is balanced by recombination that takes place as breeding
generations advance, causing it to dissipate, such that closer loci are expected to be
in higher LD than more distant ones. Consequently, the relationship between Ne and
LD affects the marker density needed to achieve and sustain adequate prediction
accuracy of a GS model across generations. In other words, marker density needs to
scale with the effective population size, and the level of LD between markers and
QTL can be increased by reducing Ne (Grattapaglia 2014).
The discussion of the extent of LD in forest trees takes us back for a moment to
the original criticism about the prospects of MAS in forest trees seen in those early
days (Strauss et al. 1992). Given the preferentially outbred nature and large Ne of
natural populations of trees, the claim was that the state of linkage equilibrium
would therefore be such that prohibitively large training populations and marker
density would be required to attain success of MAS. That original prediction was
correct in that it would apply to very conservative breeding programs that aim not
only at genetic gain but also at preserving diversity by managing breeding
populations with very large Ne > 300. However, the reality of more advanced
breeding programs where genetic gain is prioritized and alternative strategies are
devised to maintain diversity (White et al. 2007) is not one of very large effective
population sizes. Rather, small effective population sizes in the range of
Ne ¼ 10–100 are used to maximize gain. On a genome-wide basis, genetic drift is
the main contributor to LD, and drift is generated by the breeder when a closed,
selected breeding population is established. The effective population size influences
the number of independently segregating chromosome segments expected in the
population (Me) which in turn will determine the necessary genotyping density to
capture all the effects of the QTLs co-segregating with those segments. To under-
stand this relationship, a common derivation proposed for Me in populations is
Me ¼ 2NeL (Hayes et al. 2009b) where L is the genome size in Morgans. Larger Ne
and recombinationally larger genomes result in more independently segregating
chromosome segments requiring more markers. On the other hand, the smaller the
Ne, the closer the genetic relationship among individuals, the longer the indepen-
dent chromosome segments, the smaller becomes Me, and less markers are needed
to reach a certain accuracy of GS. It is important to note that the key parameter here
is not the physical genome length but rather the recombination size and number of
chromosomes. The fact that conifer genomes are very large (~20–23 Gbp) does not
matter here. Their recombination size in Morgans, ~15 M in Pinus taeda (Echt et al.
2011) to ~18 M in Picea (Pelgas et al. 2005) for 12 chromosomes, is not that
different from the recombination size of Eucalyptus, ~13 M in 11 chromosomes
(Brondani et al. 2006) with a much smaller physical genome (~0.65 Gb).
Calibrating the extent of LD by managing the effective population size, such that
near-maximum genetic gain can be achieved in a long-term breeding program, is
thus a key element when adopting GS. Theoretical studies and practical consider-
ations regarding the appropriate size of a tree breeding population have shown that
Ne between 20 and 50 will support selection with appreciable genetic gains for
several generations (Namkoong et al. 1988; White et al. 2007). Although suitable
for short-term genetic gains, such constrained Ne may be subject to larger devia-
tions of actual versus predicted progress and may result in a faster buildup of
relatedness. To remain on a conservative side in sustaining long-term gains,
effective population sizes between Ne ¼ 40 and 100 have been used, typically
corresponding to a census number (i.e., the total number of selections retained in the
breeding population in any given generation) around 200 individuals with some
level of relatedness (White et al. 2007). As examples, the third breeding cycle of
loblolly pine in the Southeastern USA has adopted a highly selected group of
40 selections to provide rapid gains (McKeand and Bridgwater 1998). In Eucalyp-
tus, populations with Ne between 30 and 60 are typically used for each species in
reciprocal recurrent selection strategies for hybrid breeding. Similar effective
population sizes are also used in recurrent selection programs based on synthetic
hybrid populations, an approach that exploits the variation derived from multiple
species aiming at the selection of elite hybrid clones for deployment (Assis and de
Resende 2011; Kerr et al. 2004). In conclusion, the effective population sizes
currently used in most tree breeding programs largely fit within the perspectives
of reaching high GS accuracies, provided that sufficient genotyping densities are
used, so that the number of independently segregating chromosome segments is
adequately tracked.
9.5.2 Genotyping Density and SNP Platforms for GS

in Forest Trees
“Recent advances in molecular genetic techniques will make dense marker maps
available and genotyping many individuals for these markers feasible.” This far-
seeing statement that introduces the seminal paper on genomic selection
(Meuwissen et al. 2001) was written when SNP discovery by Sanger sequencing
was still a prohibitively expensive endeavor for most species and SNP genotyping
platforms were in their infancy. However, it clearly recognized the key role that the
advent of faster and cheaper DNA marker genotyping would have for the new
breeding method proposed in that article. In the last 15 years, a major revolution
took place in the ability to discover large numbers of SNPs and develop new
methods to assay DNA polymorphisms, starting in 2005 with the advent of next-
generation sequencing technologies based on miniaturized and parallelized plat-
forms. What makes genomic selection different from what breeders have done so
far using the tools of quantitative genetics is the adoption of dense DNA marker
216 D. Grattapaglia
data instead of relying solely on the expected pedigree relationships. Marker data
allows one to build a genomic relationships matrix that precisely determines the
realized kinship among individuals in a breeding population. This procedure not
only allows correcting pedigree errors but, more importantly, captures the random
Mendelian sampling term resulting from gamete formation, such that the realized
genetic covariances are now based on the actual proportion of the genome that is
IBD or IBS between any two individuals (VanRaden 2008). Genomic selection
based on realized genomic relationships can produce more accurate predictions
than the pedigree-based method precisely because it exploits the variation created
by Mendelian segregation. It is therefore relevant to devote some time to discuss the
advantages and limitations of the different marker technologies currently available
for the application of GS.
Prior to the times of easy SNP discovery and genotyping, microsatellite markers
were the workhorse of genetic analysis in forest trees, and they still are for many
applications. Microsatellite genotyping has been adopted into breeding practice to
resolve clonal identity, verify parentage, and reconstruct pedigrees as a way to
reduce costs of controlled crosses (El-Kassaby and Lstiburek 2009; Grattapaglia
et al. 2004; Lambeth et al. 2001). Usually between 10 and 20 microsatellites have
been used providing abundant power to resolve parentage even with relatedness
between alleged parents. In a recent study, it was shown that some 100 selected
high-frequency SNPs are needed to match the power of 16 microsatellites for such
applications in eucalypts (Telfer et al. 2015). However, typical microsatellite
marker density is not sufficient to estimate the genome fraction shared by two
individuals and to apply this information to genomic predictions. The genotyping
density together with the effective population size showed by far the largest impact
on the prospects of GS in forest tree breeding (Grattapaglia and Resende 2011). The
upper bound benchmark accuracy of phenotypic BLUP selection, set at 0.68 in that
study, can be reached at a relatively low marker density, around 2–3 markers/cM, as
long as the effective population size is kept below Ne ¼ 60. For an average genome
of 1,500–2,000 cM, some ~5,000 SNPs would be necessary. For larger effective
population sizes up to Ne ¼ 100, however, 10 or up to 20 markers/cM would be
necessary for keeping high accuracies of GS. Such a target genotyping density will
require genotyping platforms to yield somewhere between 20,000 and 50,000
informative markers depending on the size of the recombining genome and the
effective population size of the breeding population.
The impact of the genotyping density used in the practice of GS will become
even more important as generations of selection advance. In the absence of selec-
tion, increasing marker density is beneficial to the persistence of GEBV prediction
accuracy over generations (Solberg et al. 2009) because higher marker densities
enable GEBV accuracy to persist over time due to a slower decay of LD among
tightly linked marker and trait loci. However, directional selection following the
initial training population is expected to result in a rapid decline of accuracy (Muir
2007). High-density genotyping was shown to be essential to sustain accuracy and
keep selection effective for more generations in the presence of directional selec-
tion when a finite number of QTL loci are assumed rather than an infinitesimal
model (Long et al. 2011). In such cases, selection, together with recombination,
may change the pattern of LD between markers and QTLs. The new LD generated
by selection can be unfavorable for GEBV prediction which was based on the
original marker-QTL LD structure in the training population. Although the
decrease of accuracy of GS over time can be mitigated by reestimating marker
effects or varying the weight given to markers, the possibility of using higher
genotyping densities is generally preferred.
The use of lower-density marker panels has been an interesting option to reduce
genotyping cost (Habier et al. 2009). The training population is genotyped with a
full set of markers, but selection candidates are genotyped with a smaller selected
subset. The pattern of linkage disequilibrium among markers in the training pop-
ulation is used to predict genotypes for the missing markers in the candidates. This
strategy has been successfully implemented in dairy cattle (Berry and Kearney
2011) and became a standard practice in operational GS (Boichard et al. 2012). It
could become an important strategy for GS in forest trees as well. Although theory
predicts that a lower marker density would make GS more susceptible to the decay
of LD with recombination, if prediction accuracies are mainly driven by relation-
ship, low-density marker panels would be perfectly suitable together with contin-
uous model retraining strategies. However, in forest trees the much wider genetic
diversity across breeding programs might be such that the shared use of a common
high-density SNP panels instead of each program developing a custom low-density
panel could be economically more advantageous at least in the initial stages of
GS. As GS programs of each individual organization advance and larger numbers of
samples are genotyped by each breeding program, low-density SNP panels might
become the standard practice.
9.5.2.1 Fixed Content SNP Arrays
The easy availability of shared commercial SNP chips has been a major strength of
specific communities in advancing genomic selection into operational use. The best
example of a widely uses common SNP platform for GS is the 50K bovine chip
(Matukumalli et al. 2009). Large-scale genome-wide SNP discovery projects
started relatively recently for forest tree genera such as Picea (Pavy et al. 2006),
Eucalyptus (Novaes et al. 2008), Populus (Geraldes et al. 2011), and Pinus (Eckert
et al. 2010; Lepoittevin et al. 2010). These efforts resulted in the development of
some fixed content low-density arrays with hundreds of SNPs for Pinus (Chancerel
et al. 2011; Eckert et al. 2009), Picea (Pavy et al. 2008), and Eucalyptus
(Grattapaglia et al. 2011c). Moderate-density Infinium arrays were reported for
Pinus taeda with 7,216 SNPs (Eckert et al. 2010), Picea with 9,539 SNPs (Pavy
et al. 2013), Pinus pinaster with 9,000 SNPs (Plomion et al. 2016), and higher-
density array with 34,000 SNPs for Populus (Geraldes et al. 2013). Although the
SNP contents for these arrays were published, their use was restricted to those that
developed it. They did not become commercially available products that one could
order from a vendor or buy service from a provider.
218 D. Grattapaglia
Recently, however, a high-density multispecies SNP chip for eucalypts was

developed from whole-genome resequencing of a large sample of 240 trees of
12 species (Silva-Junior et al. 2015). The EuCHIP60K, the highest-density SNP
platform so far for a forest tree, provides close to 60,000 polymorphic SNPs across
all the most widely planted and bred Eucalyptus species worldwide, providing a
96% genome-wide coverage and a density of one SNP every 12–20 kb or ~20–30
markers/cM. More importantly, however, not only its content is open source, but it
was deliberately developed as a commercial product. The EuCHIP60K was made
possible by a sort of community crowd-funding effort where eucalypt-based forest
companies mainly from Brazil agreed to genotype at least 960 trees of their
breeding programs, such that the minimum number of 15,000 samples was reached
to cover the upfront cost of chip fabrication. This SNP chip is fully available to
every interested institution, public or private, at a very competitive price through
GeneSeek (NE, USA), an agricultural genomics service provider.
Over 30,000 Eucalyptus trees have already been genotyped with the
EuCHIP60K at the time of this writing, the vast majority of the data used to start
eucalypt GS experiments and pilot programs in several forest-based companies
across the world. The use of a common genotyping platform across breeding
programs should become a very valuable asset for future research and utilization
of genomic selection. Common high-quality SNP data will allow, for example, the
development of large-scale meta-analyses of GS data opening possibilities to
develop and test prediction models based on much larger training populations.
Furthermore, genomic selection experiments carried out with this chip are now
providing the necessary information for the development of lower-density SNP
chips for specific applications.
9.5.2.2 Genotyping-by-Sequencing (GbS) Approaches
Due to the general lack of accessible SNP arrays for the majority of forest tree
species, GbS methods have been a useful entry technology to develop SNP
resources and, in some cases, carry out high-density genotyping of forest tree
populations. GbS allows capturing SNP diversity of much larger numbers of
samples and carrying out SNP discovery even in very large and complex genomes
with no reference. Large numbers of SNPs were discovered using RAD sequencing
in Eucalyptus (Grattapaglia et al. 2011a), while GbS was used to mine SNPs in
Picea glauca (Chen et al. 2013) and to carry out a genome-wide association study in
Pinus contorta (Parchman et al. 2012). Optimized GbS methods were recently
reported for three Pinus species (Pan et al. 2015). Sequence capture-based
genotyping has also been successfully applied for a more targeted complexity
reduction SNP discovery and genotyping in Populus trichocarpa (Zhou and
Holliday 2012) and Pinus taeda (Neves et al. 2014). The first study to use GbS to
carry out a GS study in forest trees was reported for Picea engelmannii (El-Dien
et al. 2015).
GbS allows simultaneous discovery and genotyping of large numbers of markers

with essentially no upfront costs (Davey et al. 2011). These methods require a
genome complexity reduction step targeting a portion of the genome for selective
enrichment, carried out either by PCR, restriction enzyme digestion, or sequence
capture, followed by high-throughput NGS to ensure high sequence coverage of the
targeted reduced representations (Cronn et al. 2012). Dominant presence/absence
variants (PAVs), derived from polymorphism in the restriction recognition sites,
and codominant single-nucleotide polymorphisms (SNPs) within the sequence tags
are detected after aligning the sequence reads with or without the aid of a reference
genome sequence. Restriction enzyme-based genome complexity reduction was
also the basic approach of the widely used Diversity Arrays Technology (DArT)
genotyping method (Kilian et al. 2012). This technology was successfully used to
develop a fixed 7,680 probe array for Eucalyptus that allowed the first reported
experimental results of genomic selection in forest trees (Grattapaglia et al. 2011b).
This highly validated hybridization-based platform was later converted to an
NGS-based assay named DArT-Seq, significantly improving throughput and
marker number for Eucalyptus (Sansaloni et al. 2011).
9.5.2.3 Fixed Content SNP Arrays Versus GbS for Genomic Selection
in Forest Trees
GbS methods have been attractive for they provide large numbers of SNPs at a
relatively lower cost per sample when compared to fixed SNP arrays. GbS does not
require prior sequence information, and there is no need to assemble a minimum
number of samples of several hundred or thousands to defray the cost of chip
fabrication. One only pays for the samples genotyped. The downside of GbS,
however, is that to keep sample costs down, the sequencing coverage is generally
low and therefore highly variable across the sampled loci in the genome (Beissinger
et al. 2013). Technical issues associated with DNA digestion, PCR amplification of
libraries, and sequencing process itself add a considerable amount of variation in
what genomic loci are sampled and at what sequence depth during sequencing. This
fact results in large proportions of missing data, usually around 40% up to 80%
depending on the depth of sequencing employed (Poland and Rife 2012). This
problem is mitigated by SNP imputation in inbred species where reference haplo-
types are easily determined by deep sequencing of founder lines and expected
genotypes are homozygous. In outbred forest trees, however, genotype imputation
is not straightforward as genomes are highly heterozygous, and multiple unrelated
parents are used such that reference haplotypes are not easily determined.
The problem of missing data in GbS tends to become substantial when
attempting to genotype complex and highly heterozygous genomes of forest trees
due to much higher restriction-site variation across individuals causing presence/
absence variants and the need of higher coverage to declare heterozygous geno-
types with confidence. This in turn leads to genotype reproducibility issues when
one attempts to genotype the same sample across independent experiments.
220 D. Grattapaglia
Genomic loci and SNPs contained into them may be sampled or not in the
replicates, and provided that the genomic locus is sampled, the genotype declared
could match or not between replicates. In a study with Poplar, out of 16 GBS
replicates of the same exact Populus trichocarpa tree Nisqually-1, the genotype
used for genome sequencing, only 27% of in silico predicted restriction sites were
sampled. Across the 16 replicates, on average, 26% of the SNPs were detected in
only one of the 16 replicates, and only 9.6% were detected in all 16 replicates. Still,
this amounted to ~34,000 loci out of the 334,000 total loci sampled. It is expected,
however, that with a larger number of samples, the proportion of SNPs genotyped
across all samples with high call rates would drop considerably. Genotype mis-
matches between replicated samples were largely due to low read coverage and
were about 2% after heavy filtering (Schilling et al. 2014). These results are in line
what has been reported for larger sample sizes genotyped by GbS in Pinus
engelmannii (El-Dien et al. 2015). Out of 1.2 million initially sampled SNPs and
after filtering by allowing 30% or up to 60% of missing data, the number of useful
SNPs dropped to 8,868 or 62,198 depending on the different imputation approaches
used. An imputation accuracy of 0.77–0.82 indicates that some ~20% of the
genotype data could still be incorrect. No mention was made of genotype repro-
ducibility in that study or the impact of such inaccuracies on genomic predictions.
It is intuitive that the success of genomic selection is dependent on SNP data
quality. One has to be able to repeatedly genotype the same set of SNPs across
generations with which the prediction models were initially developed in the
training population. It is therefore not clear at this point whether the current GbS
methods will be able to provide such data quality or, conversely, what is the
tolerable genotyping inaccuracy for successfully applying GS. Demonstration GS
proof-of-concept experiments are probably okay when carried out using GbS.
However, to implement a professional routine of long-term GS into an industrial
breeding program, very high standards of data quality should be sought. Currently
only fixed SNP array provides the gold standard of data reproducibility (>99%),
both in terms of sampling the same SNP loci and declaring the same exact genotype
for the same individual across different sample batches and laboratories. This is
probably one of the reasons why fixed SNP arrays have been the only platform used
so far in domestic animal breeding, animal model research, and human genomic
medicine.
Additionally, fixed SNP array data are breeder friendly and easily manageable
and stored without the bioinformatics burden associated with GbS data. The
common criticism of ascertainment bias of fixed content chips, a potential problem
for population genetic studies, does not represent an issue for GS (Heslot et al.
2013), reminding that any GBS method is equally subject to such bias due to the
genome complexity reduction methods involved, the biases inherent to next-
generation sequencing, and the filtering pipeline applied for data analysis. Despite
the falling prices of sequencing, the cost advantage of GbS in relation to fixed SNP
arrays has dropped substantially in recent times with more flexible chip fabrication
formats and competition among the main SNP chips vendors. It is therefore likely
that fixed arrays will also become the standard for GS in breeding of the major tree
species. Besides the EuCHIP60K already fitting such requirements, similar “com-
munity” chips are currently in development for loblolly pine (F. Isik, ‘Technical
meeting on Pine SNP chip development’ 9 September 2016). It is also true,
however, that novel, targeted GbS methods based on amplicon sequencing,
sequence capture, or padlock probes will continue to evolve in parallel to dropping
prices and increased precision of NGS platforms and analytical software. Once
quality and cost issues are carefully evaluated, the key feature in choosing a SNP
platform will be the flexibility that a new method provides to move seamlessly from
one SNP platform to another while querying the same SNP set.
9.5.2.4 Whole-Genome Sequence Data for GS
With the evolution of sequencing technologies, a discussion has taken place on the
value of moving from sparse SNP data to whole genome sequence data for the
practice of genomic selection. Notwithstanding the challenge of managing massive
NGS datasets for large numbers of individuals, in theory, if sequence data were
used instead of dense SNPs, accuracy should increase because rare causal alleles
would be better captured in the predictive models, and these in turn could be more
stable as generations advance. However, simulation studies have shown that whole-
genome sequence data does not bring any advantage in accuracy when the effective
population size is reduced and LD is longer range which is usually the case in
breeding programs (MacLeod et al. 2014). Another study found no justification to
move to whole-genome sequencing for genomic selection unless accurate prior
estimates on the functionality of SNP data could be included in the model (Perez-
Enciso et al. 2015). If all SNPs within causal genes were included in the prediction
model, accuracy could increase by ~40%. However, this advantage would be
quickly lost if incorrect or incomplete biological information regarding SNP func-
tion was used.
9.5.3 Training Population: Size, Composition,

and Phenotyping
Assembling a large number of trees into a training population to accurately estimate

SNP effects is generally not a problem in forest tree breeding, although
phenotyping costs can be an issue especially for traits that are expensive to measure.
Choice of a training population evidently will depend in large part on the breeding
strategy adopted. Training populations are typically established by sampling sev-
eral hundred or a few thousand trees in existing progeny trials at ages that will allow
extensive high-quality phenotyping of all traits targeted by the breeding program.
These trials are derived from the inter-mating (open pollinated or controlled) of a
set of a few to several dozen elite parents representative of the target germplasm,
222 D. Grattapaglia
encompassing an adequate effective population size to provide sustained gains for a

few generations ahead. Combining training sets from different populations can be
useful to boost accuracy when individual populations lack sufficient size, although
considerable risks exist of lowering the performance of such multi-population
prediction models because relatedness with the prospective selection candidates
is reduced or eliminated.
How many individuals should be included in a training population for GS in forest
tree breeding? With up to N ¼ 1,000 individuals, the selection accuracy was shown to
rapidly increase, reaching satisfactory levels. With 2,000 individuals an improvement
of ~10% in the accuracy would be expected, and larger improvements can be
achieved under conditions of lower-heritability traits, larger numbers of QTLs
involved, and larger effective population sizes. After N ¼ 2,000 simulations have
shown that the accuracy tends to plateau irrespective of the effective population size
and genotyping density (Grattapaglia and Resende 2011). However, if the QTL
distribution violates the infinitesimal model assumption of equal size effect and
common variance, not all of the genetic variance is explained, and the selection
accuracy can be lower depending on the method used to calculate the GEBV (Coster
et al. 2010). Using training sets around N ¼ 2,000 might, therefore, be warranted to
protect against such model violations or cases where several hundred QTLs control
trait variation. Simulation studies mirroring a eucalypt breeding scheme showed a
considerable improvement of genomic prediction accuracies when increasing the
training population size by consolidating phenotypic and genotypic data of individ-
uals from previous breeding cycles (Denis and Bouvet 2013). Furthermore, larger
training populations mitigate the probability of losing rare favorable alleles from the
breeding population as generations of selection advance, although some will inevi-
tably be lost because they are in low LD with any marker. A higher marker density
will also help in this respect, i.e., in preserving rarer alleles in the breeding
populations, thus allowing better long-term gains from selection.
Phenotyping large training populations of forest trees can be challenging and
expensive. To mitigate this problem, a common approach widely used in forest tree
breeding is to use indirect phenotyping methods such as NIRS (near-infrared
reflectance spectroscopy) or X-ray diffraction for high-throughput measurements
of chemical and physical wood properties. Although data collected by such
methods are generally precise, they might not be accurate to the actual whole tree
value, but they still allow confident raking of trees, which is generally satisfactory
for GS. These methods were employed in the first experimental assessments of GS
in Eucalyptus (Resende et al. 2012a) and white spruce (Beaulieu et al. 2014a, b) and
recently for the assessment of a large set of chemical and physical traits in a GS
study in Eucalyptus (Lima 2014). With current drops in genotyping costs, while
phenotyping costs remain constant or increase, considerations have been given to a
reverse approach in defining training populations so that individuals to be
phenotyped are chosen on the basis of their genotypes. For example, Rincent
et al. (2012) proposed different metrics to maximize the reliability of genomic
predictions by optimizing the composition of individuals in the training population
based exclusively on their genotypic data. Different criteria based on the diversity
or on the prediction error variance from G-BLUP prediction were proposed to select
reference individuals.
9.5.3.1 Clonal Replication of the Training Population
A common question made for tree species amenable to cloning, such as eucalypts
and poplars, is whether vegetatively propagating the individuals of a training
population would benefit the accuracy of a predictive model. In principle, by
clonally replicating individuals, trait heritability would be increased, with a likely
positive impact of accuracy. However, the impact of heritability on accuracy of GS
has been shown to be very modest when genotyping is dense (see below). To
answer this question it is relevant to remember that a key feature of GS is that
phenotyping of the training population is done to train a model, not to directly select
individuals. Selection subsequently proceeds on the basis of genomic estimated
breeding values (GEBVs) such that prediction based on allele effects is now the
selection criterion and the allele becomes the unit of evaluation. Alleles are
therefore the units that need to be replicated not individuals (Lorenz et al. 2011).
Therefore, when establishing a training population under a fixed phenotyping
budget, it is more beneficial to increase the number of individuals phenotyped
than clonally propagating and phenotyping a smaller number of individuals. Evi-
dently, however, when no budget restrictions exist as far as phenotyping, clonally
propagating a large number of individuals (N 2000) as a training population
would be advantageous, probably increasing the prediction accuracies, especially
for low-heritability traits. Additionally, clonally propagating a training population
would allow replicating it in several different environments and thus implementing
a strategy in which the same breeding population is used to breed improved genetic
material for different environments instead of a more costly option of advancing
different populations for different environments. Phenotypes collected on the same
genotypes in each environment would be used to build different prediction models
for each environment therefore optimizing budgets even in the presence of signif-
icant genotype x environment interaction.
9.5.3.2 Genetic Relationship Between Training and Selection

Candidates
The importance of relationship as a driver of accuracy in GS was shown early on

from simulation studies and underscored in all recent reviews on the perspectives
GS in plant and domestic animals breeding (Heslot et al. 2015; Lin et al. 2014; Van
Eenennaam et al. 2014). Individuals closely related to the training population are
always expected to have an advantage in accuracy over distantly related individ-
uals. The demonstration that RR-BLUP and G-BLUP are equivalent implicitly
showed that no LD between markers and QTLs is required for GS to work. The
accuracy of GEBV is nonzero even without LD. However, when SNPs are the QTL
themselves or in LD with the QTLs, RR-BLUP will provide better accuracy than
224 D. Grattapaglia
G-BLUP (Habier et al. 2007). This expectation was corroborated experimentally in

forest trees where models developed for one population had limited or no ability of
predicting phenotypes in an unrelated one (Beaulieu et al. 2014a, b; Resende et al.
2012a). These results indicate that the relatively low marker density used in these
experiments has not been able to capture LD between QTLs and markers, such that
prediction models have relied essentially on relatedness and are in principle pop-
ulation specific. Using stochastic simulations of a typical eucalypt breeding pro-
gram (Denis and Bouvet 2013) also showed a marked decrease of the prediction
accuracy at a rate of 10–15% per breeding cycle as the relationship between the
training and candidate populations decreased.
Increasing the genetic relationships between training and selection candidates
effectively has the same consequence as reducing the effective population size such
that the stronger the relationship, the higher is the accuracy. Furthermore, in maize
biparental populations it was shown that it is better to increase the accuracy of
prediction by increasing relatedness between training and validation populations,
rather than by increasing the size of the training set with less relatedness to
predicted individuals (Riedelsheimer et al. 2013). Increased relatedness reduces
the number of independently segregating chromosome segments (Me) therefore
increasing the probability that chromosome segments identical by descent sampled
in the training population are also found in the selection candidates. For the
successful implementation of GS it is therefore crucial that the selection candidates
are genetically related to the training population.
The issue of relationship is one that should also be carefully considered when cross
validating prediction models. The individuals on whom the models will be applied
are the selection candidates, but the accuracy of predicting their phenotypes cannot be
estimated because their phenotypes are not available. The models are therefore tested
by cross-validation, typically using a subsample of the training population. Because
relatedness is an important component of prediction accuracy, the most important
principle of selecting a testing population is that it should mirror the relationship of
the selection candidates to the training population (Daetwyler et al. 2013). If the
testing population is more or less related to the training population than the selection
candidates, then the prediction accuracy will be over- or underestimated, respec-
tively. In replicated cross-validation, the manner in which individuals are assigned to
particular folds affects accuracy. Random assignment of individuals to training or
testing sets is prone to inflate accuracies because of within-family components
driving them. A more realistic approach is to randomly assign whole full- or half-
sib families to training or testing sets to evaluate prediction accuracy across families
or to design cross-validation schemes that use genomic relationship data to partition
individuals into the various folds to minimize the relationships between training and
testing populations (Saatchi et al. 2011).
Beaulieu et al. (2014a) carried out the most informative study so far to evaluate
the impact of genetic relationship on the accuracy of genomic prediction in outbred
forest trees. A training population of 1,694 trees representative of 214 open-
pollinated families was phenotyped for 12 wood and growth traits and genotyped
for 6,385 SNPs. Three cross-validation schemes were applied with decreasing
relationship between training and validation sets. CV1 involved allowing half-sib
relationships between the sets, CV2 was performed by eliminating all maternal
relatedness by assigning entire families to folds of the training and validation sets,
and CV3 was designed to control for any possible contribution of the pollen parent
to relatedness, thus eliminating as much as possible any possibility of coancestry
between training and validation sets. Confirming expectations, they found that
predictive ability between remotely related individuals (CV2) was only slightly
lower (5–20% depending on the trait) than that of those built for closely related
individuals (CV1). When the possibility of coancestry between cross-validation
sets was eliminated and confirmed by an average estimated kinship coefficient of
zero, the prediction accuracy was considerably reduced but still clearly different
from zero for several of the traits supporting the putative presence of historical LD
between SNPs and trait loci, despite the relatively sparse SNP data. In a subsequent
study (Beaulieu et al. 2014b), this time dealing with two totally unrelated breeding
groups, good predictions were obtained within each breeding group. However a
sharp drop of accuracies near zero was seen when training was carried out in one
group and cross-validated in the other. SNP genotyping was low density (6,932
SNPs) for an estimated recombining genome of 2,100 cM, and SNPs on the chip
were not evenly distributed across the genome but rather targeted a limited set of
candidate genes. The ability to capture historical LD was therefore very limited, if
any, further confirming the key role of relationship as the pivotal driver of accuracy
in these genomic prediction experiments.
Understanding the drivers of prediction accuracy in GS is a relevant issue
because it has a direct impact on the ability of GS models to predict phenotypes
in future generation removed from training. The concept of GS, as originally
outlined, was based on the understanding that LD alone would explain the predic-
tive ability of a model (Meuwissen et al. 2001). Later, however, it became clear
both from simulation and experimental studies that prediction accuracy was also
affected by genetic relationships between training and validation sets captured by
SNP data (Habier et al. 2007; Legarra et al. 2008). The relative contributions of LD,
additive genetic relationship, and cosegregation to the accuracy of predictions were
modeled under different scenarios and their persistence over generations assessed
(Habier et al. 2013). Among the several results that those simulations revealed, it
was shown that the correlation between GEBVs within families depends largely on
additive genetic relationship, which is determined by the size of the training
populations and the effective number of SNPs. This latter one was defined as the
number of ideal SNPs that provides the same accuracy due to additive genetic
relationships as the actual number of SNPs in the model. It decreases with the
increasing range of LD and therefore with decreasing effective population size,
explaining why the accuracy due to additive genetic relationships does not improve
beyond a certain SNP density. The lack of improvement in accuracy with increasing
number of SNPs or, conversely, the rapid attainment of a plateau of accuracy with
relatively few hundred SNPs has been a common observation in almost all GS
studies in forest trees to date (Beaulieu et al. 2014a, b; Resende et al. 2012a, b).
Surprisingly, the same prediction accuracies were seen whether using SNPs
226 D. Grattapaglia
selected based on highest estimated effects for any particular trait or chosen
randomly (Beaulieu et al. 2014b) providing strong evidence in support of relation-
ships as the main source of accuracy. While the LD component of predictive
accuracy is expected to persist over generations without the need for retraining,
the component due to additive genetic relationships is anticipated to decay rapidly
with the successive generations of recombination.
9.5.3.3 Populations for GS in Hybrid Breeding
Breeding for interspecific hybrids is an established strategy in some of the main

plantation forest tree species. Hybrids combine desirable traits from two or more
species through complementation of additive gene action. The best documented
example in trees is the E. grandis x E. urophylla hybrid that combines growth and
fungal disease resistance, respectively, and displays a heterotic effect due to
nonadditive gene action. Elite hybrid individuals are clonally deployed, frequently
exhibiting greater phenotypic stability that allows extending plantation range to
sites where one or both parental species have a suboptimal performance (Rezende
et al. 2014). Exploiting hybrid breeding in eucalypts can be done in a single
synthetic population where the original species are hybridized at the outset to
form a single breeding population which is then advanced by conventional recur-
rent selection (Kerr et al. 2004). Alternatively, reciprocal recurrent selection
between the two species can be adopted, and the breeding goal in the pure species
is to optimize the performance of hybrid descendants that are deployed either as
clones or hybrid seed varieties.
The question arises on what would be the population to train a model for hybrid
breeding under such a reciprocal recurrent selection strategy. Would it be the hybrid
population or the pure species populations? The maintenance of predictive ability
of a GS model across different populations or species will essentially rely on the
consistency of LD across them, which in turn depends on the recombination rate
between marker and QTLs and the time since the two diverged. The less diverged
the populations are and the higher the marker density, better performance of the
predictive model is expected across populations. An analogous situation takes place
in bovine breeding in which selection is carried out in pure breeds, but the aim is to
improve crossbred performance. Results from simulation studies generally show
that training on crossbred data provides good prediction accuracy for selecting
purebred individuals for crossbred performance. The incorporation of dominance in
the model and the use of high-marker densities are generally beneficial (de Roos
et al. 2009; Ibanz-Escriche et al. 2009; Kizilkaya et al. 2010; Zeng et al. 2013).
When crossbred data is not available, separate purebred training populations can be
used either separately or combined depending on the correlation of LD phase
between the pure lines (Esfandyari et al. 2015).
In trying to make a parallel between the bovine breeding scenario and the case of
eucalypt hybrid breeding, a prediction model could be trained on a hybrid popula-
tion, i.e., a hybrid progeny trial, and used to select individuals in the two pure
species for their performance as parents of hybrids. It is important to note, however,

that while the estimates of the age of the most recent common ancestor of domes-
ticated cattle range from 200 to 300 KYA (Murray et al. 2010), the estimated
divergence time between the Eucalyptus species used in hybrid breeding is much
older at 2–5 MYA (Silva-Junior and Grattapaglia 2015) such that it is not clear at
this point if such an approach would actually work given the much wider diver-
gence. If such a strategy proves effective, however, the hybrid progeny trial would
also serve to train a model to select hybrid candidates to be deployed as clones. In
other words, a prediction model involving additive effects would in principle be
developed to select for high GEBV individuals in each species separately to serve
as parents of the subsequent generation. If nonadditive effects are relevant to the
target traits, a separate model including also nonadditive components would better
serve to select individuals based on their genomic estimated genotypic value
(GEGV) for clonal deployment. A simulation study showed that a GS model
including dominance effect outperforms an additive model only when the training
population is large and updated by combination of data from previous breeding
cycles (Denis and Bouvet 2013). Clearly, experimental examination of the potential
approaches and feasibility of applying GS to reciprocal recurrent selection in hybrid
eucalypt populations deserve further attention.
9.5.4 Trait Heritability and Genetic Architecture
Theory predicts that the number of QTLs underlying trait variation will have an
important impact on the accuracy of GS. Fewer loci controlling larger fractions of
the phenotypic variance are more easily captured relative to a more complex
genetic architecture involving larger numbers of loci with smaller effects. As
pointed out earlier, QTL mapping experiments in forest trees have revealed increas-
ing numbers of QTLs controlling each trait as more and larger mapping populations
were used. It is reasonable, therefore, to assume that quantitative traits are con-
trolled by several tens to hundreds of QTLs. Simulations have shown that the
reduction of GS accuracy with an increasing number of QTLs tends to be more
pronounced at lower-marker densities or larger effective population sizes. Assum-
ing a total of 200 QTLs, marker densities 5–10 markers/cM would be necessary
assuming a simpler genetic architecture, while 20 markers/cM would be necessary
with larger numbers of QTLs (Grattapaglia and Resende 2011).
Heritability on the other hand was shown to have a relatively minor impact on
accuracy when the training population size is large enough so that marker effects
are adequately estimated. GS accuracy is directly proportional to the product of the
heritability and the ratio between the number of phenotypic records in the training
population and the number of QTLs involved (Daetwyler et al. 2008). Therefore,
by simulating a scenario with a rather modest training set for a tree breeding
situation of N 1000 individuals, a trait controlled by 100 QTL, and an effective
population size Ne ¼ 60, the GS accuracy increased only slightly, from 0.71 to 0.83,
as the heritability went from 0.2 to 0.6 (Grattapaglia and Resende 2011).
228 D. Grattapaglia
Simulation studies for animal breeding scenarios also showed that a decrease in
accuracy with decreasing heritability is readily compensated by using larger train-
ing sets (Meuwissen et al. 2001; Nielsen et al. 2009).
9.5.5 Data Analysis Approaches for GS in Forest Trees
Genomic prediction requires methods that are capable of handling cases where the
number of marker variables ( p) greatly exceeds the number of individuals (n) (the
large p small n problem) while mitigating the risk of model over parameterization.
Several analytical approaches have been proposed and used for prediction of
genome-estimated breeding or genotypic values. Ideally, a genomic prediction
method should provide high accuracy, limit over-fitting on the training population,
and preferably capture marker-QTL LD rather than relatedness for higher long-term
stability. A good method should be easy to implement, reliable across a wide range
of traits and datasets, and computationally efficient (Heslot et al. 2012). Several
thorough reviews are available regarding the features of the main prediction
methods for GS (de los Campos et al. 2013; Heffner et al. 2009; Lorenz et al.
2011), guidelines to compare them (Daetwyler et al. 2013), and comparative
benchmark assessments in animal (Moser et al. 2009), crops (Heslot et al. 2012),
and forest trees (Resende et al. 2012c). The current methods basically differ with
respect to the assumptions regarding the genetic architecture of the trait for which
genomic predictions are sought.
For the scope of this discussion, it is relevant to highlight the fact that across
several reports in crops, trees, and domestic animals, the ridge regression best linear
unbiased prediction (RR-BLUP) method using a mixed model has been very effective
in providing the best compromise between computation time and prediction effi-
ciency (Lorenz et al. 2011). RR-BLUP assumes that the trait is controlled by many
loci of small effect, so that all marker effects are treated as random, normally
distributed, and with a common variance. Results therefore suggest that most eco-
nomically important quantitative traits adequately fit into the assumption of an
infinitesimal model. In a loblolly pine study, for example, the performance of
RR-BLUP and three Bayesian methods was only marginally different when com-
pared across 17 traits with distinct heritabilities, with a small improvement using
BayesA only for fusiform rust resistance where loci of relatively larger effect had
been described (Resende et al. 2012c). Equivalent results were obtained for growth
and wood traits in other forest trees showing no performance difference between
RR-BLUP and Bayesian methods (Beaulieu et al. 2014b; Isik et al. 2016; Lima 2014;
Ratcliffe et al. 2015). Considering the overall efficiency of RR-BLUP or the equiv-
alent G-BLUP, a general recommendation has been made to use it as a starting point
from which to explore additional alternative models (Heslot et al. 2012; Lorenz et al.
2011), although additional research in this area is warranted. Additional models
would include Bayesian methods, when suspicion or prior information exists re-
garding the existence of loci of larger effect, or machine learning methods when
nonadditive effects are known or presumed important.
9.5.5.1 Modeling Nonadditive Effects in Genomic Prediction
In several plant species, and particularly in some forest trees such as eucalypts,
vegetative propagation of outstanding individuals is a key strategy for deploying
elite genetic material. Clones maximize gains from selection by capturing additive
and nonadditive effects. In forest trees, it is also common to observe that top parents
may not be top clones and vice versa, suggesting considerable levels of nonadditive
variation depending on the trait. A dominance to additive variance ratio close to 1.2
for growth was estimated in E. grandis x E. urophylla (Bouvet et al. 2009), while in
E. globulus this ratio was 0.8 with indications that epistasis might be the main
component of the nonadditive variance (Araujo et al. 2012). GS for tree breeding
has therefore received increased attention in evaluating models including nonaddi-
tive effects. A simulation study directed to Eucalyptus breeding showed that a
model including dominance effects performed better for clone selection only when
dominance effects were preponderant (i.e., a dominance to additive variance ratio
approaching 1.0) and heritability was >0.6 (Denis and Bouvet 2013).
Genomic data has also been successfully used to understand the relative impor-
tance of additive versus nonadditive variation and its implication in tree breeding. A
number of studies have shown that the accuracy and stability of prediction models
were improved by using marker-based instead of pedigree-based relationship
matrices (Beaulieu et al. 2014a; Bouvet et al. 2016; Munoz et al. 2014; Zapata-
Valenzuela et al. 2013). Besides correcting pedigree errors, marker-based matrices
capture both the Mendelian segregation within full-sib families and genetic links
through unknown common ancestors which are not available in the known pedi-
gree. In Pinus taeda the use of a genomic relationships matrix yielded a better
separation of additive and nonadditive components of the variance in height growth
when compared to the pedigree-based model. Results provided evidence that
additive pedigree-based models tend to inflate breeding values by capturing a
large proportion of variance due to interaction terms. Additionally, it was shown
that models including nonadditive relationship were more stable than traditional
G-BLUP at predicting breeding values (Munoz et al. 2014). In hybrid eucalypts,
using genome-wide information was also shown to improve the variance partition
(Bouvet et al. 2016). At this point, however, no experimental data exist yet in forest
trees regarding the ability of GS in predicting the total genotypic value of individual
trees including additive and nonadditive effects, across generations. Research into
this topic is one of the top priorities for forest tree species that are deployed as
clonal varieties.
9.5.5.2 Genomic Prediction as a Ranking Problem
When judging the potential value of genomic prediction for selection, it is essential
that the training and validation scheme adopted must reflect the way genomic
prediction will be used in practice. The discussion on the feasibility of selecting
individual trees for clonal propagation takes us to the recognition that until now, the
predictive accuracy of a model has been typically assessed using the Pearson
230 D. Grattapaglia
correlation between the observed trait values and the predicted trait values. GS has
been essentially formulated as a regression problem. However, in tree breeding
programs where clonal propagation is possible and clones are a genetic “dead end,”
i.e., they are not used back in breeding, it is common that the unit of selection and
deployment is the individual tree. The breeder is simply interested in ranking
individuals for their own merit from the best to the worst, without necessarily
predicting their breeding value. This is particularly relevant to hybrid breeding
strategies in eucalypts where individual trees are selected, ranked, tested, and
eventually deployed as clones. When evaluating what individual tree ranking
would GS reveal as compared to standard BLUP phenotypic selection, Lima
(2014) reported a coincidence above 70% when selecting the top 30 trees out of
the 1,000 of the training population by leave-one-out cross-validation and of 60%
when tandem selection for volume growth and wood density was applied.
In a recent study, particularly relevant to tree breeding strategies that target
individual tree selection for clonal propagation, Blondel et al. (2015) proposed to
formulate GS as a ranking problem, showing that Pearson’s correlation may
correlate poorly with individual ranking accuracy. The approach also involves
model estimation and candidate selection stages. However, instead of imposing
that the model satisfies the equivalence of predicted with observed value, a score is
assigned to each candidate, and the scores are used to rank the candidates. Machine
learning methods were employed to rank individuals in six different datasets of both
inbred and outbred plants. The approach showed a significantly higher efficiency to
correctly rank individuals when compared to several standard regression methods.
Clearly, this study opens a new avenue of GS research to develop methods that
better fit the case of selecting top individuals for clonal propagation.
9.5.6 Genomic Prediction Accuracy Across Environments

and Ages
G*E is essentially a lack of consistency in the relative performance of individuals

when they are grown in different environments. Genotype by environment (G*E)
interaction is a fact that all tree breeding programs commonly deal with. G*E can
be of different levels, depending on the species, environmental variability and
extent of the intended forest plantation sites, and type of planting material,
whether families or clones, with clones typically being more interactive than
families. Interactions can be more subtle when differences in performance are
observed, but the relative ranking of tested individuals does not change across
different environments (termed scale-effect interaction) or more severe types of
interactions when rank changes are observed. Correct ranking of individual trees
by their genetic value is a key component of the successful implementation of
GS. Therefore, while the presence of scale-effect interactions should not represent
a major limitation of a prediction model, rank changes are critical. When large
rank change interactions are found, the GS strategy must account for this
(Grattapaglia 2014).
Considerations and treatment of the interaction between genome predictions and
environment will follow the same procedures used in dealing with standard G*E
effects. Technically, there is nothing different between dealing with conventional
G*E or genomic effect by environment interaction. The same consideration regard-
ing the definition of breeding or management zones (i.e., the set of environments for
which an improved variety is being developed), commonly applied to tree breeding
programs, will apply to GS as well. Prediction models might be accurate across
sites within the same breeding zone but probably not across breeding zones.
However, the need to develop specific GS models for each breeding zone will
largely depend upon the type of interaction observed, whether scale effect or rank
change.
In forest trees multi-environmental G*E interaction not only is commonplace,
but it is used to assess the performance of the same clones or families across
different environmental conditions, to study genotype stability and to predict the
performance of untested genotypes. Heffner et al. (2009) pointed out that GS opens
the opportunity to evaluate the effect of particular genomic segments that are shared
between lines across multiple environments. This information sharing should
provide GS with stability of predictions even in the presence of G*E. This concept
was put in practice by Burgueno et al. (2012) using a multi-environment dataset of
wheat lines, showing that combining pedigree and marker data can yield substantial
increases in prediction accuracy relative to traditional pedigree-based prediction
and to single-environment pedigree and genomic prediction models. Multi-
environment GS models enhanced predictive power in across-environment predic-
tion, i.e., predicting the performance of genotypes that were evaluated in some
environments but not in others.
The impact of environmental variation on the success of genomic predictions in
forest trees has been evaluated, corroborating the expectations based on previous
knowledge of G*E trends. Generally, all studies showed an important impact of
environmental variation, but its magnitude and variation across traits followed what
was already known from G*E studies, with growth traits showing higher interaction
than wood properties. Resende et al. (2012b) clonally propagated and deployed the
same set of 951 loblolly pine individuals in four locations on a north-south gradient
along the Southeastern USA. Prediction models trained using the local phenotypes
provided good predictions within site, but predictions got increasingly poorer as the
geographical distance between training and testing sites increased along the latitu-
dinal gradient. In white spruce, across-environment predictions were essentially the
same as those within environment for wood traits but dropped for growth traits,
confirming the contrasting behavior previously seen for these traits in typical G*E
studies (Beaulieu et al. 2014b).
The accuracy of GS models in predicting the GEBV was assessed in interior
spruce (Picea glauca x P. engelmannii) using a set of 1,126 38-year-old trees
planted across three different sites originating from 25 open-pollinated families.
Predictions of seven growth and wood traits were evaluated using four
232 D. Grattapaglia
cross-validation scenarios: (1) training and validation within each individual site;
(2) cross-site validation (all possible combinations); (3) within multisite, i.e., the
three sites combined into a single training set; and (4) multisite training and
validation in each individual site (El-Dien et al. 2015). Good accuracies were
obtained when training and testing were carried out within each site despite the
small training population available within each site, but, as expected, they dropped
across sites. The estimated type-b genetic correlations between sites closely
reflected the trend observed for the GS accuracy observed across sites. Prediction
accuracies of a single multisite training model were higher for all seven traits when
compared to the accuracies estimated in within-site validations, likely driven by the
considerably larger training population used in this scenario with all 1,126 trees.
Similarly, when the multisite model was validated on each separate site, accuracies
were essentially as good as within site, suggesting that the positive effect of
increasing the training population size counterbalanced the effect of environmental
variation.
Another key aspect in forest tree breeding is the impact of age on the accuracy of
predictions. Ideally, selection should be applied on trees at the same age when they
are usually harvested. However, it is common for tree breeders to make selections at
an earlier age in an attempt to accelerate a breeding program. The feasibility of such
an approach will depend essentially on the magnitude of the age-age or juvenile-
mature correlations which can be relatively high for wood quality traits but low for
growth traits, although ample variation exists depending on species, environment,
and ages considered (White et al. 2007). GS accuracy across ages was assessed in
loblolly pine using diameter and height growth measurements obtained over mul-
tiple years (Resende et al. 2012b). As expected, given the weak juvenile-mature
correlations typically observed in conifers (Namkoong et al. 1988), GS models
trained on phenotypes measured at ages 1–2 years had unacceptable accuracy in
predicting phenotypes at age 6 years. Equivalent results were reported in a recent
study in Picea engelmannii using a series of repeated tree height measurements
through ages 3–40 years on a population of 769 trees belonging to 25 open-
pollinated families. Prediction accuracies varied substantially through time
mirroring the spatial competition among trees. As expected, the behavior of geno-
mic prediction accuracies across time was highly correlated with age-age genetic
correlations and decreased substantially with increasing difference in age between
the training and validation populations (Ratcliffe et al. 2015).
Results of the experimental studies reported to date on the impact of environ-
ment and age on the accuracy of genomic prediction in forest trees lead to a general
conclusion. Existing data from traditional G*E or age-age correlation studies will
inform with good precision what to expect from genomic prediction across envi-
ronments and ages. As a rule, accurate predictions will require training models on
traits measured at the same age and environment as the ones where predictions on
selection candidates are planned. As age-age correlations between training and
testing age improve and the magnitude and trend of the G*E interaction becomes
inconsequential between training and testing sites, predictions will tend to be
satisfactory, provided that genetic relationship between training and selection

candidates is kept in the population.
There is an additional aspect to be considered about the prospects of GS across
environments that was examined in the context of GS in crops (Heslot et al. 2015),
but that will potentially be much more relevant and challenging in forest trees due
to their longer life span. We have seen above that the appropriate tree phenotypes to
train a prediction model should be collected at or very close to harvest age, which
usually spans several years or decades, and preferably in the same target environ-
ment as the one where GS will be practiced on future selection candidates. The
target environment of a forest tree is a consolidation of the action of several abiotic
and biotic factors during the long and variable preceding life of the tree, including
severe droughts, frosts, pest, and diseases attacks. Assuring that the target environ-
ment where phenotypes are collected for model training will be the same for the
future selection candidates may therefore be much more critical (and challenging)
for GS than in conventional phenotypic selection. In the latter, phenotypic data are
used only to rank and select individuals on which phenotypes were measured. Thus,
if a particular year of data is a misleading sample of the target environment due to
some severe climate fluctuations during the life span of the tree, it will impact
genetic gain for only that particular generation of selection. In GS, on the other
hand, the unrepresentative data may affect genetic gain over a much longer period
of time, as it will influence marker effect estimates that, in turn, will affect selection
criteria going several generations forward. Periodical retraining with phenotypes
collected in more recent generations of breeding might help mitigate this problem.
Finally, despite the challenges of dealing with G*E, GS provides opportunities to
integrate environmental covariates (e.g., climate data) to predict G*E deviations for
unobserved environments. This approach can in turn allow prediction of individual
stability, identification of important stresses, and understanding of the target envi-
ronmental variation that is critical for breeding strategies (Heslot et al. 2015;
Jarquin et al. 2014).
9.5.7 Performance of GS Across Generations
Proof-of-concept experiments in forest trees have been carried out by sampling

training and validation sets within the same generation, usually the same progeny
trial or different progeny trials, involving the same set of half- or full-sib families.
Marker density was generally low, with a few thousand markers only, and accuracy
was mostly driven by relatedness and not by marker-QTL LD. Not only experi-
mental data is still lacking on simple two-generation cross-validation, but nothing is
known about the performance of GS for long-term gain. The duly posed question by
breeders is how accurate will the genomic predictions be on individuals several
generations removed from the training population? As generations advance, recom-
bination will erode both marker-QTL LD and links of relatedness between training
and selection candidates reducing accuracy, while directional selection may change
234 D. Grattapaglia
both the genetic architecture of the trait, via changes in allele frequencies, and the
patterns of LD making them potentially unfavorable for GEBV prediction.
Recently, the first GS studies to evaluate GS accuracy by training and validating
models using individuals of three generations G0, G1, and G2 were carried out in
maritime pine (Pinus pinaster). In a first study, 184 individuals of the G0 parental
and 477 of the G1 progeny generation were used (Isik et al. 2016). Mixed sets of
parents and progeny were used for training and validation, resulting in good
predictive abilities (0.43–0.49) for stem sweep, total height, and tree diameter.
Recently, that population was expanded by including individuals of a G2 popula-
tion. G2 individuals were preselected to include exclusively individuals that would
limit the effective population size to Ne ¼ 25, fully confirmed pedigree and highest
BLUP for volume and stem straightness. Following simulations to select the best
subsample of G2 individuals to maximize prediction accuracy, models trained on
46 G0 and 62 G1 individuals and validated on 710 G2 individuals showed high
(0.70–0.85) predictive abilities despite the very small training population size,
possibly a result boosted by the preselection of G2 individuals maximizing relat-
edness. Therefore, while promising results of GS have been reported in essentially
all forest tree studies (Table 9.1), strictly speaking only one result of genomic
prediction across generations is available so far, although several experiments are in
the ground as we speak. Despite the inherent limitations of GS models validated
exclusively within generation, they could still be quite useful in situations where the
same crosses are repeated and prediction is applied on sibs of the original training
set to increase selection intensity. This approach would be particularly useful to
select top individuals to be deployed as clones by capturing additive and nonaddi-
tive effects, especially for late-expressing traits. However, when GS is applied to
advance generations, selection candidates will rarely belong to the same population
as the training set and may well be several generations removed from it.
Experimental studies assessing the performance of GS across multiple genera-
tions of breeding take some time to happen or rely on existing individuals of
ancestral generations like the Pinus pinaster described above. However, several
studies approached this issue by simulations. In the seminal study of Meuwissen
et al. (2001), the decline of GS accuracy over generations was estimated at 5% per
generation, getting smaller in later generations. Other studies under more complex
models including the effect of directional selection and the structure and depth of
the training population have been reported (Bastiaansen et al. 2012; Iwata et al.
2011; Jannink 2010; Long et al. 2011; Muir 2007; Sonesson and Meuwissen 2009).
All these studies fundamentally converged to a similar recommendation: marker
effects have to be reestimated frequently in order to maintain accuracy of pre-
dictions over generations. The issue of model updating was specifically assessed for
a 60-year conifer tree breeding program by comparing the performance of GS with
conventional phenotypic selection using stochastic simulations (Iwata et al. 2011).
Results showed that GS outperformed phenotypic selection in the short term
(30 years) but not in the long term (60 years). When the prediction model was
updated, however, the genetic gain of GS was nearly twice that of phenotypic
selection, even for low-heritability traits, with a greater advantage of GS as
genotyping density increased. Two model updating strategies were tested. In a more
conventional one, the prediction model generated in the initial cycle of selection is
updated after three (or more) generations of GS by carrying out a progeny trial of
already genotyped selection candidates, and their data is used to reestimate the
marker effects. In a second strategy, in each cycle of GS, a subset of the genotyped
selection candidates of that cycle is planted in a progeny trial. After a few years
(depending on the species), phenotypes for that subset of trees become available
and are used to update the prediction model. From that point on, every year the
prediction model gets updated with the inclusion of phenotypic data of the extra
subset of trees from previous generations. Because a set of trees from every cycle of
GS is actually field grown, this second updating strategy allows continuous verifi-
cation of the genetic progress of the GS program, although it involves greater costs
of growing and measuring trees every generation and could theoretically increase
the probability of unintended fixation of unfavorable alleles (Iwata et al. 2011). A
significant advantage of model updating on GS accuracy by including phenotypic
data from previous cycles was also shown by simulations in the context of Euca-
lyptus breeding (Denis and Bouvet 2013).
From the practical standpoint of a breeding program, continuously associating
phenotypic data from previous cycles of GS and thus progressively updating
prediction models and increasing the size and pedigree depth of the training
population seem to be a very sensible and feasible approach to adopt. The cost of
genotyping the subsets of selection candidates would have already been covered in
the GS cycle, and growing and measuring a few hundred trees would not represent a
significant cost while allowing for permanent monitoring of the realized perfor-
mance of GS. Such a continuous retraining approach would allow the additive
relationship component of predictive ability to be sustained across generations such
that GS could be successfully practiced despite a limited ability to capture
SNP-QTL LD due to the lower genotyping densities necessary to keep costs
affordable in a breeding program.
9.5.8 Inbreeding and Maintenance of Genetic Diversity

with GS
Finally, two additional issues have been raised regarding the performance of GS
over the long term: inbreeding and loss of useful variation. GS could potentially
result in a fast and unintended frequency increase of deleterious alleles causing
inbreeding depression or fixation of unfavorable QTL alleles due to the progressive
effect of drift with the restriction of effective population size. Daetwyler et al.
(2007) showed that GS reduces the rate of inbreeding per generation when com-
pared with sib and BLUP selection. High accuracies of estimated breeding values
are achieved through better prediction of the Mendelian sampling term. This
genomic-level resolution increases differentiation among sibs, allowing the breeder
236 D. Grattapaglia
to better manage coancestry and to mitigate the rate of inbreeding even when
selecting related individuals in breeding programs that are pushing for high genetic
gains. Consistent with this expectation, the effect of nonrandom mating on the rate
of inbreeding was found to be smaller for breeding schemes that adopt genome
predictions when compared to conventional mating and selection designs (Nirea
et al. 2012).
While GS is more efficient in reducing pedigree-based inbreeding when com-
pared to BLUP by increasing emphasis on the individual rather than family
information, pedigree inbreeding might not accurately reflect loss of genetic vari-
ation and the true level of inbreeding due to changes in allele frequencies and hitch-
hiking. Liu et al. (2014) evaluated this issue using simulations, concluding that GS
can have a greater impact than pedigree-based BLUP on the reduction of genetic
diversity surrounding QTLs by a “hitch-hiking” effect because GS leads to a higher
accuracy of selection on the QTL. Another reason might be that instead of directly
selecting the QTL, selection acts on markers in LD with the QTL. This effect
becomes more important when QTL effects are large, such that when implementing
long-term genomic selection, genomic control of inbreeding is therefore essential to
reduce the considerable hitch-hiking effects that are associated with genomic
selection, regardless of the prediction model used.
The second issue regarding the impact of GS over time relates to the loss of
favorable alleles with the faster successive cycles of breeding, potentially causing a
progressive reduction of response to selection. Besides the loss of useful diversity,
the hitch-hiking effect could also increase the frequency of linked deleterious
alleles. Measures to mitigate this effect include using higher genotyping densities,
periodical model updating, and verification of performance of a subset of selected
trees along the GS cycles of breeding to monitor any possible reduction of vigor
attributable to weakly or moderately deleterious mutations (Iwata et al. 2011).
Additionally it has been shown that adopting weighed GS (Goddard 2009) together
with using a larger training set (Jannink 2010) will help reducing the loss of
low-frequency favorable alleles in the breeding population, although some will
inevitably be lost due to low LD with any genotyped marker. In a simulation study,
Jannink (2010) showed that placing additional weight on low-frequency favorable
marker alleles allowed GS to increase their frequency earlier on, causing an initial
increase in genetic variance. This procedure led to higher long-term gain while
mitigating losses in short-term gain. Weighted GS also increased the maintenance
of marker polymorphism, ensuring that QTL-marker linkage disequilibrium was
higher than in conventional unweighted GS.
9.6 Conclusions and Perspectives
A number of recent experimental reports have now showed that the prospects of GS
applied to forest tree breeding are very encouraging. To illustrate how one would
envisage the operational flow of GS in tree breeding, Fig. 9.2 outlines the
Fig. 9.2 Comparative timelines of genomic selection (GS) breeding and phenotypic selection
(PS) breeding for tropical Eucalyptus (see text for details) (Modified from Grattapaglia (2014))
comparative timelines of breeding by GS and breeding by standard phenotypic

selection for a recurrent selection strategy in tropical Eucalyptus. Both methods
start at year zero with the same breeding population, and in the GS route, it is
assumed that predictive models were previously developed. In a GS breeding cycle,
following SNP genotyping and genomic prediction of all target traits (i.e., growth,
238 D. Grattapaglia
form, wood properties, disease resistance, etc.), selection candidates can follow
three possible nonexclusive routes:
1. Top ranked seedlings for GEBV (genomic estimated breeding value) are imme-
diately routed to flower induction treatment and recombined to create the
improved population (green boxes) completing the recurrent breeding cycle
(green boxes);
2. Top ranked seedlings for GEGV (genomic estimated genotypic value) are
cloned directly by mini-cutting methods and deployed in field verification clonal
trials and ultimately submitted to a final selection for elite operational clones for
plantation (red boxes);
3. A random subset of a few hundred selection candidates in each GS cycle can be
planted in field trials to provide in due course additional phenotypic data to be
added to the initial training dataset allowing continuous predictive model
updating (gray boxes).
In the proposed scheme, GS is expected to eliminate the field progeny trial

phase, accelerating the completion of a breeding cycle by allowing the selection of
elite clones much faster. With GS, a cycle of recurrent selection in tropical
Eucalyptus breeding, going from an original population to an improved population,
will last 5 years, while in standard breeding it lasts at least 10 years. Two gener-
ations of elite clones can be developed by GS in 14 years, while standard pheno-
typic selection will only provide one generation in 15 years. Note that in standard
breeding the verification clonal trial lasts 5–6 years to allow adequate phenotyping
of wood properties traits. In GS, although accurate predictions of wood properties
traits should obtained by GEBV, still this 5–6 years verification clonal trial is kept
mostly to validate the general field performance and adaptability of the prospective
clones. Depending on the performance of GS as the program proceeds, it might
eventually be possible to preclude or shorten this final verification clonal trial,
therefore further accelerating the deployment of new clones into the commercial
forest.
The effective application of genomic prediction in a tree breeding program will
vary on a case-by-case basis following a detailed cost-benefit analysis. GS might
not be an option for small-scale breeding programs for tree species with a limited or
niche market share, little prior genetic information on the species, and modest
budgets. On the other hand, for aggressive breeding programs of the major tree
species that support large industrial forest-based operations, it seems clear that time
gains by eliminating progeny testing and streamlining clonal trials of young
genomically ranked trees for multiple traits should be valuable. The adoption of
GS might therefore become a competitive advantage in turning breeding genera-
tions quicker and thus deploying improved genetic stocks in the commercial forest
at a faster rate. In concluding this chapter, it seems therefore useful to review the
main lessons learned that have emerged so far from the experimental reports of
genomic prediction in forest trees. They are summarized in a nine-point tentative

roadmap that should assist tree breeders and managers when considering research
or operational implementation of GS in their organizations:
1. Starting Population for Training GS Models. Leveraging existing progeny trials
of the current breeding program, consisting of several tens of half- or full-sib
families with relatively constrained effective population sizes, has been a suc-
cessful approach to establish training populations. By mirroring actual tree
breeding settings, the satisfactory prediction abilities estimated in essentially
all studies and for all traits are good indications of the promising operational
prospects of GS. Sampling preferably 2,000 individuals or at least 1,000 from
such progeny trials for detailed phenotyping and SNP genotyping should be an
effective way to establish a robust training population to start a GS program.
2. Genetic Relationship Between Training Population and Selection Candidates.
For the successful implementation of GS, it is crucial that the selection candi-
dates are genetically related to the training population. Studies that evaluated the
impact of removing relatedness between training and validation sets have pro-
vided strong evidence in this respect. Higher genotyping densities and evalua-
tion across multiple generations of breeding will now be needed to assess the
relative importance of the decay of relatedness versus the SNP-QTL LD in
maintaining satisfactory prediction abilities. Model updating strategies will
likely be very important to counteract the expected decay of relationship and
LD such that good prediction abilities might still be maintained with relatively
sparse SNP genotyping densities.
3. Genotyping Platform. Studies in forest trees have shown satisfactory predictive
abilities using relatively modest genotyping densities (2,500 ~ 10,000 SNPs)
likely due to the leading role of relatedness as driver of accuracy. Higher marker
densities should however be recommended to capture true LD and sustain long-
term accuracies. There is ample room for improvement of SNP genotyping
platforms in parallel with the development and experimental assessment of
lower-density marker panels. While improved genotyping-by-sequencing
(GbS) methods will likely surface in the near future, at this point fixed SNP
array technologies unquestionably constitute the gold standard for data quality
and breeder friendliness. Costs of such arrays have dropped significantly in
recent years, although they still require upfront development costs which can
be easily shared by interested organizations. This has been successfully done for
species of Eucalyptus where a public SNP chip is available. Similar efforts are
underway for species of Pinus such that high-standard public SNP genotyping
platforms are today realistic targets for the mainstream plantation forest tree
species.
4. Genotype by Environment and Age Interactions. Studies that evaluated the
impact of G*E on the efficiency of GS showed that predictive abilities were
reduced when models trained in one environment were validated in a different
one, although the magnitude of such reduction varied across traits. Data from
240 D. Grattapaglia
traditional G*E or age-age correlation studies will inform with good precision
what to expect from genomic prediction across environments and ages. As a
general rule, accurate predictions will require training models on traits measured
at the same age and environment as the ones where predictions on selection
candidates are planned. As age-age correlations between training and testing age
improve and the magnitude and trend of the G*E interaction becomes inconse-
quential between training and testing sites, predictions will tend to be satisfac-
tory, provided that genetic relationships between training and selection
candidates are kept in the population.
5. Data Analysis. A genomic prediction method should provide high accuracy,
limit over-fitting on the training population, and capture marker-QTL LD
besides relatedness for higher long-term stability. A good method should be
easy to implement, reliable across a wide range of traits and datasets, and
computationally efficient. Across several reports in crops, trees, and domestic
animals, the RR-BLUP method and the G-BLUP equivalent have been effective
in providing the best compromise between computation time and prediction
efficiency. RR-BLUP assumes that the trait is controlled by many loci of small
effect, therefore suggesting that most economically important quantitative traits
in forest trees adequately fit into the assumption of an infinitesimal model.
Several open-access softwares are available to implement this method, and
training courses on their use are regularly offered by several institutions world-
wide. Still, research on the subject is warranted to develop improved approaches
including methods to efficiently incorporate nonadditive variation and individ-
ual ranking of trees for clonal selection.
6. Logistics. Logistic issues such as specific nursery infrastructure, sample collec-
tion and tracking system, large-scale DNA extraction and qualification,
genotyping service providers, and data analysis pipelines are equally important
modules for the successful implementation of a GS operation but beyond the
scope of this chapter. Nevertheless, several of these components are either
already routinely used in standard nursery operations of large forest-based
companies or can be easily established in-house (e.g., DNA extraction lab) or
accessed through specialized service providers in agricultural genomics.
7. Cost-Benefit Analysis. A detailed cost-benefit analysis of adopting GS using net
present value methodologies is an absolutely necessary step before considering
its implementation. The groundbreaking advance that GS caused in dairy cattle
breeding is frequently used as an example of the economically successful use of
this technology. It has been questioned whether it is an adequate benchmark for
annual crops, although not so for forest trees where GS was considered to be
potentially even more successful than in dairy cattle (Jonas and de Koning
2013). Still, while cattle and trees share the same challenge of long generation
times, the logistics and cost of progeny testing a bull is substantially higher than
progeny testing a tree, such that the cost of genotyping is easily justified and a
remarkable gain in selection intensity has been possible. Current cost of
genotyping a sample even at USD ~40 is still expensive for many forest tree
breeding programs and more so for those that run on very tight budgets.
Assembling very large numbers of samples across breeding programs of several
organizations on the same SNP genotyping array in long-term contracts is
expected, however, to provide the necessary economy of scale to drive costs
down in the near future.
8. Prediction Across Generations. Despite the encouraging estimates of predictive
ability so far, it should be stressed once again that all studies but a recent one
only evaluated the potential of GS within the same generation. In other words,
training and validation sets were contemporary. Results of GS across indepen-
dent generations of parents and progeny are limited so far and much less the
performance of GS in generations farther removed from training. Moreover, the
impact of recombination and selection across generations on prediction ability
could not yet be assessed too, an issue that might become more relevant as
generations of GS advance. There is a general urgency among research groups
working in the area to provide additional experimental data on actual genomic
selection across generations. Several experiments are underway especially in
eucalypts, to compare the ranking of individual trees predicted at seedling stage
based on genomic data with their realized ranking at rotation age for growth and
wood quality traits.
9. Changing Environment and Model Retraining. Assuring that the target environ-
ment where phenotypes are collected for model training will be the same for the
future selection candidates is a challenging issue for GS. In conventional
breeding, phenotypic data are used only to rank and select individuals on
which phenotypes were measured. Thus, if a particular year of data is a mis-
leading sample of the target environment, it will impact genetic gain for only a
short period of time. In GS, on the other hand, unrepresentative phenotypic data
collected in the training population will affect genetic gain over a much longer
period of time, affecting selection criteria going forward. Periodical retraining
with phenotypes collected in more recent generations of breeding that were
exposed to more recent environments should mitigate this problem.
GS is definitely a hot topic in tree breeding and a fast-moving area of research in
several organizations worldwide, both public and private, working on the interface
of genomics and quantitative genetics. While some of the fundamental genetic
aspects discussed here are not likely to change much, or are valid under current
technologies and circumstances, some others will almost unquestionably change in
the future as new genotyping and sequencing technologies materialize and
improved statistical approaches are developed. As GS adoption evolves and large
experimental datasets are gathered across unrelated populations of tens of thou-
sands of trees, the accumulation of genomic prediction data should also provide a
powerful experimental framework, beyond QTL mapping and association genetics,
toward the fundamental investigation of complex trait variation. The evolution of
integrative approaches based on such large genotype and phenotype datasets should
242 D. Grattapaglia
deliver important additional hints toward understanding the connections and inter-
actions between the multitude of discrete genome-wide elements and the continu-
ous phenotypic variation in complex traits. The full elucidation of such connections
will nevertheless continue to be a very challenging endeavor due to the time and
space dynamics of the effects of these genomic elements and the stochastic pro-
cesses that thwart the expected one-to-one relationship between genotypes and
phenotypes.
Acknowledgments This work was supported by CNPq grant 577047/2008-6, PRONEX-FAP-DF

grant “NEXTREE” 2009/00106-8, EMBRAPA Macroprogram 2 grant 02.07.01.004, and a CNPq
research fellowship to DG. Special thanks to all my students, collaborators, and colleagues
worldwide working in genomic prediction and forest tree breeding with whom I have had the
privilege to share and discuss several of the ideas presented in this chapter.
References
Araujo JA, Borralho NMG, Dehon G (2012) The importance and type of non-additive genetic
effects for growth in Eucalyptus globulus. Tree Genet Genomes 8:327–337
Arus P, Verde I, Sosinski B, Zhebentyayeva T, Abbott AG (2012) The peach genome. Tree Genet
Genomes 8:531–547
Assis TF, de Resende MDV (2011) Genetic improvement of forest tree species. Crop Breed Appl
Biotechnol 11:44–49
Bartholome J, Van Heerwaarden J, Isik F, Boury C, Vidal M, Plomion C, Bouffier L (2016)
Performance of genomic prediction within and across generations in maritime pine. BMC
Genomics 17:604
Bastiaansen JWM, Coster A, Calus MPL, van Arendonk JAM, Bovenhuis H (2012) Long-term
response to genomic selection: effects of estimation method and reference population structure
for different genetic architectures. Genet Sel Evol 44:3
Beaulieu J, Doerksen T, Clement S, Mackay J, Bousquet J (2014a) Accuracy of genomic selection
models in a large population of open-pollinated families in white spruce. Heredity
113:343–352
Beaulieu J, Doerksen TK, MacKay J, Rainville A, Bousquet J (2014b) Genomic selection
accuracies within and between environments and small breeding groups in white spruce.
BMC Genomics 15:1048
Beissinger TM, Hirsch CN, Sekhon RS, Foerster JM, Johnson JM, Muttoni G, Vaillancourt B,
Buell CR, Kaeppler SM, de Leon N (2013) Marker density and read depth for genotyping
populations using genotyping-by-sequencing. Genetics 193:1073–1081
Bernardo R (2008) Molecular markers and selection for complex traits in plants: learning from the
last 20 years. Crop Sci 48:1649–1664
Bernardo R, Yu JM (2007) Prospects for genome wide selection for quantitative traits in maize.
Crop Sci 47:1082–1090
Berry DP, Kearney JF (2011) Imputation of genotypes from low- to high-density genotyping
platforms and implications for genomic selection. Animal 5:1162–1169
Blondel M, Onogi A, Iwata H, Ueda N (2015) A ranking approach to genomic selection. PLoS One
10:e0128570
Boichard D, Chung H, Dassonneville R, David X, Eggen A, Fritz S, Gietzen KJ, Hayes BJ, Lawley
CT, Sonstegard TS, Van Tassell CP, PM VR, Viaud-Martinez KA, Wiggans GR, Consortium
BL (2012) Design of a bovine low-density SNP array optimized for imputation. PLoS One 7:
e34130
Bouvet JM, Makouanzi G, Cros D, Vigneron P (2016) Modeling additive and non-additive effects
in a hybrid population using genome-wide genotyping: prediction accuracy implications.
Heredity 116:146–157
Bouvet JM, Saya A, Vigneron P (2009) Trends in additive, dominance and environmental effects
with age for growth traits in Eucalyptus hybrid populations. Euphytica 165:35–54
Brondani RP, Williams ER, Brondani C, Grattapaglia D (2006) A microsatellite-based consensus
linkage map for species of Eucalyptus and a novel set of 230 microsatellite markers for the
genus. BMC Plant Biol 6:20
Burgueno J, de los Campos G, Weigel K, Crossa J (2012) Genomic prediction of breeding values
when modeling genotype x environment interaction using pedigree and dense molecular
markers. Crop Sci 52:707–719
Chancerel E, Lepoittevin C, Le Provost G, Lin YC, Jaramillo-Correa JP, Eckert AJ, Wegrzyn JL,
Zelenika D, Boland A, Frigerio JM, Chaumeil P, Garnier-Gere P, Boury C, Grivet D,
Gonzalez-Martinez SC, Rouze P, Van de Peer Y, Neale DB, Cervera MT, Kremer A, Plomion
C (2011) Development and implementation of a highly-multiplexed SNP array for genetic
mapping in maritime pine and comparative mapping with loblolly pine. BMC Genomics
12:368
Charlier C, Coppieters W, Rollin F, Desmecht D, Agerholm JS, Cambisano N, Carta E, Dardano S,
Dive M, Fasquelle C, Frennet JC, Hanset R, Hubin X, Jorgensen C, Karim L, Kent M,
Harvey K, Pearce BR, Simon P, Tama N, Nie H, Vandeputte S, Lien S, Longeri M,
Fredholm M, Harvey RJ, Georges M (2008) Highly effective SNP-based association mapping
and management of recessive defects in livestock. Nat Genet 40:449–454
Chen C, Mitchell SE, Elshire RJ, Buckler ES, El-Kassaby YA (2013) Mining conifers’ mega-
genome using rapid and efficient multiplexed high-throughput genotyping-by-sequencing
(GBS) SNP discovery platform. Tree Genet Genomes 9:1537–1544
Coster A, Bastiaansen JWM, Calus MPL, van Arendonk JAM, Bovenhuis H (2010) Sensitivity of
methods for estimating breeding values using genetic markers to the number of QTL and
distribution of QTL variance. Genet Sel Evol 42:9
Cronn R, Knaus BJ, Liston A, Maughan PJ, Parks M, Syring JV, Udall J (2012) Targeted
enrichment strategies for next-generation plant biology. Am J Bot 99:291–311
Crossa J, de los Campos G, Perez P, Gianola D, Burgueno J, Araus JL, Makumbi D, Singh RP,
Dreisigacker S, Yan JB, Arief V, Banziger M, Braun HJ (2010) Prediction of genetic values of
quantitative traits in plant breeding using pedigree and molecular markers. Genetics
186:713–724
Daetwyler HD, Calus MPL, Pong-Wong R, de los Campos G, Hickey JM (2013) Genomic
prediction in animals and plants: simulation of data, validation, reporting, and benchmarking.
Daetwyler HD, Villanueva B, Bijma P, Woolliams JA (2007) Inbreeding in genome-wide selec-
tion. J Anim Breed Genet 124:369–376
Davey JW, Hohenlohe PA, Etter PD, Boone JQ, Catchen JM, Blaxter ML (2011) Genome-wide
genetic marker discovery and genotyping using next-generation sequencing. Nat Rev Genet
12:499–510
de los Campos G, Hickey JM, Pong-Wong R, Daetwyler HD, Calus MPL (2013) Whole-genome
regression and prediction methods applied to plant and animal breeding. Genetics 193:327–345
de Roos APW, Hayes BJ, Goddard ME (2009) Reliability of genomic predictions across multiple
populations. Genetics 183:1545–1553
Denis M, Bouvet JM (2013) Efficiency of genomic selection with models including dominance
effect in the context of Eucalyptus breeding. Tree Genet Genomes 9:37–51
244 D. Grattapaglia
Dillen S, Storme V, Marron N, Bastien C, Neyrinck S, Steenackers M, Ceulemans R, Boerjan W

(2008) Genomic regions involved in productivity of two interspecific poplar families in
Europe. 1. Stem height, circumference and volume. Tree Genet Genomes 5:147–164
Echt CS, Saha S, Krutovsky KV, Wimalanathan K, Erpelding JE, Liang C, Nelson CD (2011) An
annotated genetic map of loblolly pine based on microsatellite and cDNA markers. BMC Genet
12:17
Eckert AJ, Pande B, Ersoz ES, Wright MH, Rashbrook VK, Nicolet CM, Neale DB (2009) High-
throughput genotyping and mapping of single nucleotide polymorphisms in loblolly pine
(Pinus taeda L.) Tree Genet Genomes 5:225–234
Eckert AJ, van Heerwaarden J, Wegrzyn JL, Nelson CD, Ross-Ibarra J, Gonzalez-Martinez SC,
Neale DB (2010) Patterns of population structure and environmental associations to aridity
across the range of loblolly pine (Pinus taeda l., Pinaceae). Genetics 185:969–982
El-Dien OG, Ratcliffe B, Klapste J, Chen C, Porth I, El-Kassaby YA (2015) Prediction accuracies
for growth and wood attributes of interior spruce in space using genotyping-by-sequencing.
BMC Genomics 16:370
El-Kassaby YA, Lstiburek M (2009) Breeding without breeding. Genet Res 91:111–120
Esfandyari H, Sorensen AC, Bijma P (2015) Maximizing crossbred performance through purebred
genomic selection. Genet Sel Evol 47:16
Freeman JS, Potts BM, Downes GM, Pilbeam D, Thavamanikumar S, Vaillancourt RE (2013)
Stability of quantitative trait loci for growth and wood properties across multiple pedigrees and
environments in Eucalyptus globulus. New Phytol 198:1121–1134
Geraldes A, Difazio SP, Slavov GT, Ranjan P, Muchero W, Hannemann J, Gunter LE, Wymore
AM, Grassa CJ, Farzaneh N, Porth I, Mckown AD, Skyba O, Li E, Fujita M, Klapste J,
Martin J, Schackwitz W, Pennacchio C, Rokhsar D, Friedmann MC, Wasteneys GO, Guy RD,
El-Kassaby YA, Mansfield SD, Cronk QCB, Ehlting J, Douglas CJ, Tuskan GA (2013) A 34K
SNP genotyping array for Populus trichocarpa: design, application to the study of natural
populations and transferability to other Populus species. Mol Ecol Resour 13:306–323
Geraldes A, Pang J, Thiessen N, Cezard T, Moore R, Zhao YJ, Tam A, Wang SC, Friedmann M,
Birol I, Jones SJM, Cronk QCB, Douglas CJ (2011) SNP discovery in black cottonwood
(Populus trichocarpa) by population transcriptome resequencing. Mol Ecol Resour 11:81–92
Gion JM, Carouche A, Deweer S, Bedon F, Pichavant F, Charpentier JP, Bailleres H, Rozenberg P,
Carocha V, Ognouabi N, Verhaegen D, Grima-Pettenati J, Vigneron P, Plomion C (2011)
Comprehensive genetic dissection of wood properties in a widely-grown tropical tree: Euca-
lyptus. BMC Genomics 12:301
Goddard M (2009) Genomic selection: prediction of accuracy and maximisation of long term
response. Genetica 136:245–257
Goddard ME, Hayes BJ (2009) Mapping genes for complex traits in domestic animals and their use
in breeding programmes. Nat Rev Genet 10:381–391
Grattapaglia D (2014) Breeding forest trees by genomic selection: current progress and the way
forward. Chapter 26. In: Tuberosa R, Graner A, Frison E (eds) Advances in genomics of plant
genetic resources. Springer, New York, pp 652–682
Grattapaglia D, Chaparro J, Wilcox P, Mccord S, Werner D, Amerson H, Mckeand S,
Bridgwater F, Whetten R, O’malley D, Sederoff RR (1992) Mapping in woody plants with
RAPD markers: applications to breeding in forestry and horticulture. Proceedings of the
symposium “applications of RAPD technology to plant breeding”. Crop Science Society of
America, American Society of Horticultural Science, American Genetic Association, pp 37–40
Grattapaglia D, de Alencar S, Pappas G (2011a) Genome-wide genotyping and SNP discovery by
ultra-deep Restriction-Associated DNA (RAD) tag sequencing of pooled samples of E. grandis
and E. globulus. BMC Proc 5:P45
Grattapaglia D, Plomion C, Kirst M, Sederoff RR (2009) Genomics of growth traits in forest trees.
Curr Opin Plant Biol 12:148–156
Grattapaglia D, Resende MDV (2011) Genomic selection in forest tree breeding. Tree Genet
Genomes 7:241–255
Grattapaglia D, Resende MDV, Resende M, Sansaloni C, Petroli C, Missiaggia A, Takahashi E,
Zamprogno K, Kilian A (2011b) Genomic selection for growth traits in Eucalyptus: accuracy
within and across breeding populations. BMC Proc 5:O16
Grattapaglia D, Ribeiro VJ, Rezende GD (2004) Retrospective selection of elite parent trees using
paternity testing with microsatellite markers: an alternative short term breeding tactic for
Eucalyptus. Theor Appl Genet 109:192–199
Grattapaglia D, Sederoff R (1994) Genetic-linkage maps of Eucalyptus-grandis and Eucalyptus-
urophylla using a pseudo-testcross – mapping strategy and RAPD markers. Genetics
137:1121–1137
Grattapaglia D, Silva OB, Kirst M, de Lima BM, Faria DA, Pappas GJ (2011c) High-throughput
SNP genotyping in the highly heterozygous genome of Eucalyptus: assay success, polymor-
phism and transferability across species. BMC Plant Biol 11:65
Grattapaglia D, Vaillancourt R, Shepherd M, Thumma B, Foley W, Külheim C, Potts B, Myburg A
(2012) Progress in Myrtaceae genetics and genomics: Eucalyptus as the pivotal genus. Tree
Genet Genomes 3:463–508
Greenwood MS, Adams GW, Gillespie M (1991) Stimulation of flowering by grafted black spruce
and white spruce – a comparative-study of the effects of gibberellin A4/7, cultural treatments,
and environment. Can J For Res 21:395–400
genome-assisted breeding values. Genetics 177:2389–2397
Habier D, Fernando RL, Dekkers JCM (2009) Genomic selection using low-density marker panels.
Genetics 182(1):343–353
Habier D, Fernando RL, Garrick DJ (2013) Genomic BLUP decoded: a look into the black box of
genomic prediction. Genetics 194:597–607
Haley CS, Visscher PM (1998) Strategies to utilize marker-quantitative trait loci associations. J
Dairy Sci 81:85–97
Harfouche A, Meilan R, Kirst M, Morgante M, Boerjan W, Sabatti M, Mugnozza GS (2012)
Accelerating the domestication of forest trees in a changing world. Trends Plant Sci 17:64–72
Hasan O, Reid JB (1995) Reduction of generation time in Eucalyptus-globulus. Plant Growth
Regul 17:53–60
Hayes BJ, Bowman PJ, Chamberlain AJ, Goddard ME (2009a) Invited review: genomic selection
in dairy cattle: progress and challenges. J Dairy Sci 92:433–443
Hayes BJ, Visscher PM, Goddard ME (2009b) Increased accuracy of artificial selection by using
the realized relationship matrix. Genet Res 91:47–60
Heffner EL, Sorrells ME, Jannink JL (2009) Genomic selection for crop improvement. Crop Sci
49:1–12
Heslot N, Jannink JL, Sorrells ME (2015) Perspectives for genomic selection applications and
research in plants. Crop Sci 55:1–12
Heslot N, Rutkoski J, Poland J, Jannink JL, Sorrells ME (2013) Impact of marker ascertainment
bias on genomic selection accuracy and estimates of genetic diversity. PLoS One 8:e74612
Heslot N, Yang HP, Sorrells ME, Jannink JL (2012) Genomic selection in plant breeding: a
comparison of models. Crop Sci 52:146–160
Ibanz-Escriche N, Fernando RL, Toosi A, Dekkers JCM (2009) Genomic selection of purebreds
for crossbred performance. Genet Sel Evol 41:12
Ingvarsson PK, Garcia MV, Luquez V, Hall D, Jansson S (2008) Nucleotide polymorphism and
phenotypic associations within and around the phytochrome B2 Locus in European aspen
(Populus tremula, Salicaceae). Genetics 178:2217–2226
Isik F, Bartholome J, Farjat A, Chancerel E, Raffin A, Sanchez L, Plomion C, Bouffier L (2016)
Genomic selection in maritime pine. Plant Sci 242:108–119
Iwata H, Hayashi T, Tsumura Y (2011) Prospects for genomic selection in conifer breeding: a
simulation study of Cryptomeria japonica. Tree Genet Genomes 7:747–758
246 D. Grattapaglia
Jannink JL (2010) Dynamics of long-term genomic selection. Genet Sel Evol 42:35
Jannink JL, Zhong SQ, Dekkers JCM, Fernando RL (2009) Factors affecting accuracy from
genomic selection in populations derived from multiple inbred lines: a barley case study.
Jarquin D, Crossa J, Lacaze X, Du Cheyron P, Daucourt J, Lorgeou J, Piraux F, Guerreiro L,
Perez P, Calus M, Burgueno J, de los Campos G (2014) A reaction norm model for genomic
selection using high-dimensional genomic and environmental data. Theor Appl Genet
127:595–607
tion accuracy. Genetics 192:1513–1522
Jonas E, de Koning DJ (2013) Does genomic selection have a future in plant breeding? Trends
Biotechnol 31:497–504
Jonas E, de Koning DJ (2015) Genomic selection needs to be carefully assessed to meet specific
requirements in livestock breeding programs. Front Genet 6:49
Junghans DT, Alfenas AC, Brommonschenkel SH, Oda S, Mello EJ, Grattapaglia D (2003)
Resistance to rust (Puccinia psidii Winter) in eucalyptus: mode of inheritance and mapping
of a major gene with RAPD markers. Theor Appl Genet 108:175–180
Kerr RJ, Dieters MJ, Tier B (2004) Simulation of the comparative gains from four different hybrid
tree breeding strategies. Can J For Res 34:209–220
Kilian A, Wenzl P, Huttner E, Carling J, Xia L, Blois H, Caig V, Heller-Uszynska K, Jaccoud D,
Hopper C, Aschenbrenner-Kilian M, Evers M, Peng K, Cayla C, Hok P, Uszynski G (2012)
Diversity arrays technology: a generic genome profiling technology on open platforms.
Methods Mol Biol 888:67–89
Kizilkaya K, Fernando RL, Garrick DJ (2010) Genomic prediction of simulated multibreed and
purebred performance using observed fifty thousand single nucleotide polymorphism geno-
types. J Anim Sci 88:544–551
Lambeth C, Lee BC, O’Malley D, Wheeler NC (2001) Polymix breeding with parental analysis of
progeny: an alternative to full-sib breeding and testing. Theor Appl Genet 103:930–943
Legarra A, Robert-Granie C, Manfredi E, Elsen JM (2008) Performance of genomic selection in
mice. Genetics 180:611–618
Lepoittevin C, Frigerio JM, Garnier-Gere P, Salin F, Cervera MT, Vornam B, Harvengt L,
Plomion C (2010) In vitro vs in silico detected SNPs for the development of a genotyping
array: what can we learn from a non-model species? PLoS One 5:e11034
Lima BM (2014) Bridging genomics and quantitative genetics of Eucalyptus: genome-wide
prediction and genetic parameter estimation for growth and wood properties using high-
density SNP data. Genetics Dep. University of S~ao Paulo, Piracicaba, SP, Brazil, pp 93.
Available in English at http://www.teses.usp.br/teses/disponiveis/11/11137/tde-25062014-
25085814/pt-br.php
Lin Z, Hayes BJ, Daetwyler HD (2014) Genomic selection in crops, trees and forages: a review.
Crop Pasture Sci 65:1177–1191
Liu HM, Sorensen AC, Meuwissen THE, Berg P (2014) Allele frequency changes due to hitch-
hiking in genomic selection programs. Genet Sel Evol 46:8
Long N, Gianola D, Rosa GJM, Weigel KA (2011) Long-term impacts of genome-enabled
selection. J Appl Genet 52:467–480
Lorenz AJ, Chao SM, Asoro FG, Heffner EL, Hayashi T, Iwata H, Smith KP, Sorrells ME, Jannink
JL (2011) Genomic selection in plant breeding: knowledge and prospects. Adv Agron
110:77–123
MacLeod IM, Hayes BJ, Goddard ME (2014) The Effects of demography and long-term selection
on the accuracy of genomic prediction with sequence data. Genetics 198 (4):1671–1684
Matukumalli LK, Lawley CT, Schnabel RD, Taylor JF, Allan MF, Heaton MP, O’Connell J,
Moore SS, Smith TPL, Sonstegard TS, Van Tassell CP (2009) Development and characteri-
zation of a high density SNP genotyping assay for cattle. PLoS One 4:e5350
McKeand SE, Bridgwater FE (1998) A strategy for the third breeding cycle of loblolly pine in the
Southeastern US. Silvae Genet 47:223–234
Moser G, Tier B, Crump RE, Khatkar MS, Raadsma HW (2009) A comparison of five methods to
predict genomic breeding values of dairy bulls from genome-wide SNP markers. Genet Sel
Evol 41:56
Muir WM (2007) Comparison of genomic and traditional BLUP-estimated breeding value accu-
racy and selection response under alternative trait and genomic parameters. J Anim Breed
Genet 124:342–355
Munoz PR, Resende MFR, Gezan SA, Resende MDV, de los Campos G, Kirst M, Huber D, Peter
GF (2014) Unraveling additive from nonadditive effects using genomic relationship matrices.
Genetics 198:1759–1768
Murray C, Huerta-Sanchez E, Casey F, Bradley DG (2010) Cattle demographic history modelled
from autosomal sequence variation. Philos T R Soc B 365:2531–2539
Namkoong G, Kang HC, Brouard JS (1988) Tree breeding: principles and strategies. Springer
Verlag, New York
Neale DB, Kremer A (2011) Forest tree genomics: growing resources and applications. Nat Rev
Genet 12:111–122
Neale DB, Williams CG (1991) Restriction-fragment-length-polymorphism mapping in conifers
and applications to forest genetics and tree improvement. Can J For Res 21:545–554
Nejati-Javaremi A, Smith C, Gibson JP (1997) Effect of total allelic relationship on accuracy of
evaluation and response to selection. J Anim Sci 75:1738–1745
Neves LG, Davis JM, Barbazuk WB, Kirst M (2014) A high-density gene map of loblolly pine
(Pinus taeda L.) based on exome sequence capture genotyping. G3 Genes Genom Genet
4:29–37
Nielsen HM, Sonesson AK, Yazdi H, Meuwissen THE (2009) Comparison of accuracy of genome-
wide and BLUP breeding value estimates in sib based aquaculture breeding schemes. Aqua-
culture 289:259–264
Nirea KG, Sonesson AK, Woolliams JA, Meuwissen THE (2012) Effect of non-random mating on
genomic and BLUP selection schemes. Genet Sel Evol 44:11
Novaes E, Drost DR, Farmerie WG, Pappas GJ Jr, Grattapaglia D, Sederoff RR, Kirst M (2008)
High-throughput gene and SNP discovery in Eucalyptus grandis, an uncharacterized genome.
BMC Genomics 9:312
Novaes E, Osorio L, Drost DR, Miles BL, Boaventura-Novaes CRD, Benedict C, Dervinis C,
Yu Q, Sykes R, Davis M, Martin TA, Peter GF, Kirst M (2009) Quantitative genetic analysis of
biomass and wood chemistry of Populus under different nitrogen levels. New Phytol
182:878–890
Pan J, Wang BS, Pei ZY, Zhao W, Gao J, Mao JF, Wang XR (2015) Optimization of the
genotyping-by-sequencing strategy for population genomic analysis in conifers. Mol Ecol
Resour 15:711–722
Parchman TL, Gompert Z, Mudge J, Schilkey FD, Benkman CW, Buerkle CA (2012) Genome-
wide association genetics of an adaptive trait in lodgepole pine. Mol Ecol 21:2991–3005
Pavy N, Gagnon F, Rigault P, Blais S, Deschenes A, Boyle B, Pelgas B, Deslauriers M, Clement S,
Lavigne P, Lamothe M, Cooke JEK, Jaramillo-Correa JP, Beaulieu J, Isabel N, Mackay J,
Bousquet J (2013) Development of high-density SNP genotyping arrays for white spruce
(Picea glauca) and transferability to subtropical and nordic congeners. Mol Ecol Resour
13:324–336
Pavy N, Parsons LS, Paule C, MacKay J, Bousquet J (2006) Automated SNP detection from a large
collection of white spruce expressed sequences: contributing factors and approaches for the
categorization of SNPs. BMC Genomics 7:174
Pavy N, Pelgas B, Beauseigle S, Blais S, Gagnon F, Gosselin I, Lamothe M, Isabel N, Bousquet J
(2008) Enhancing genetic mapping of complex genomes through the design of
248 D. Grattapaglia
highly-multiplexed SNP arrays: application to the large and unsequenced genomes of white
spruce and black spruce. BMC Genomics 9:1–17
Pelgas B, Bousquet J, Beauseigle S, Isabel N (2005) A composite linkage map from two crosses
for the species complex Picea mariana x Picea rubens and analysis of synteny with other
Pinaceae. Theor Appl Genet 111:1466–1488
Perez-Enciso M, Rincon JC, Legarra A (2015) Sequence- vs. chip-assisted genomic selection:
accurate biological information is advised. Genet Sel Evol 47:43
Plomion C, Bartholome J, Lesur I, Boury C, Rodriguez-Quilon I, Lagraulet H, Ehrenmann F,
Bouffier L, Gion JM, Grivet D, de Miguel M, de Maria N, Cervera MT, Bagnoli F, Isik F,
Vendramin GG, Gonzalez-Martinez SC (2016) High-density SNP assay development for
genetic analysis in maritime pine (Pinus pinaster). Mol Ecol Resour 16:574–587
Poland JA, Rife TW (2012) Genotyping-by-sequencing for plant breeding and genetics. Plant
Genome 5:92–102
Pryce JE, Daetwyler HD (2012) Designing dairy cattle breeding schemes under genomic selection:
a review of international research. Anim Prod Sci 52:107–114
Rae A, Pinel M, Bastien C, Sabatti M, Street N, Tucker J, Dixon C, Marron N, Dillen S, Taylor G
(2008) QTL for yield in bioenergy Populus: identifying GE interactions from growth at three
contrasting sites. Tree Genet Genomes 4:97–112
Ratcliffe B, El-Dien OG, Klapste J, Porth I, Chen C, Jaquish B, El-Kassaby YA (2015) A
comparison of genomic selection models across time in interior spruce (Picea engelmannii x
glauca) using unordered SNP imputation methods. Heredity 115:547–555
Resende MDV, Resende MFR, Sansaloni CP, Petroli CD, Missiaggia AA, Aguiar AM, Abad JM,
Takahashi EK, Rosado AM, Faria DA, Pappas GJ, Kilian A, Grattapaglia D (2012a) Genomic
selection for growth and wood quality in Eucalyptus: capturing the missing heritability and
accelerating breeding for complex traits in forest trees. New Phytol 194:116–128
Resende MFR, Munoz P, Acosta JJ, Peter GF, Davis JM, Grattapaglia D, Resende MDV, Kirst M
(2012b) Accelerating the domestication of trees using genomic selection: accuracy of predic-
tion models across ages and environments. New Phytol 193:617–624
Resende MFR, Munoz P, Resende MDV, Garrick DJ, Fernando RL, Davis JM, Jokela EJ, Martin
TA, Peter GF, Kirst M (2012c) Accuracy of genomic selection methods in a standard data set of
loblolly pine (Pinus taeda L.) Genetics 190:1503–1510
Rezende GDSP, Resende MDV, Assis TF (2014) Eucalyptus breeding for clonal forestry. In:
Fenning T (ed) Challenges and opportunities for the world’s forests in the 21st century.
Springer ScienceþBusiness Media, Dordrecht, pp 393–424
Riedelsheimer C, Endelman JB, Stange M, Sorrells ME, Jannink JL, Melchinger AE (2013)
Genomic predictability of interconnected biparental maize populations. Genetics 194:493–503
Rincent R, Laloe D, Nicolas S, Altmann T, Brunel D, Revilla P, Rodriguez VM, Moreno-
Gonzalez J, Melchinger A, Bauer E, Schoen CC, Meyer N, Giauffret C, Bauland C, Jamin P,
Laborde J, Monod H, Flament P, Charcosset A, Moreau L (2012) Maximizing the reliability of
genomic selection by optimizing the calibration set of reference individuals: comparison of
methods in two diverse groups of maize inbreds (Zea mays L.) Genetics 192:715–728
Saatchi M, McClure MC, McKay SD, Rolf MM, Kim J, Decker JE, Taxis TM, Chapple RH,
Ramey HR, Northcutt SL, Bauck S, Woodward B, Dekkers JCM, Fernando RL, Schnabel RD,
Garrick DJ, Taylor JF (2011) Accuracies of genomic breeding values in American Angus beef
cattle using K-means clustering for cross-validation. Genet Sel Evol 43:40
Sansaloni C, Petroli C, Jaccoud D, Carling J, Detering F, Grattapaglia D, Kilian A (2011) Diversity
Arrays Technology (DArT) and next-generation sequencing combined: genome-wide, high
throughput, highly informative genotyping for molecular breeding of Eucalyptus. BMC Proc 5:
P54
Schilling MP, Wolf PG, Duffy AM, Rai HS, Rowe CA, Richardson BA, Mock KE (2014)
Genotyping-by-sequencing for populus population genomics: an assessment of genome
sampling patterns and filtering approaches. PLoS One 9:95292
Silva-Junior OB, Faria DA, Grattapaglia D (2015) A flexible multi-species genome-wide 60K SNP
chip developed from pooled resequencing 240 Eucalyptus tree genomes across 12 species.
New Phytol 206:1527–1540
Silva-Junior OB, Grattapaglia D (2015) Genome-wide patterns of recombination, linkage disequi-
librium and nucleotide diversity from pooled resequencing and single nucleotide polymor-
phism genotyping unlock the evolutionary history of Eucalyptus grandis. New Phytol
208:830–845
Solberg TR, Sonesson AK, Woolliams JA, Odegard J, Meuwissen THE (2009) Persistence of
accuracy of genome-wide breeding values over generations when including a polygenic effect.
Genetics Selection Evolution 41 (1):53
Sonesson AK, Meuwissen THE (2009) Testing strategies for genomic selection in aquaculture
breeding programs. Genet Sel Evol 41:37
Stirling B, Newcombe G, Vrebalov J, Bosdet I, Bradshaw HD (2001) Suppressed recombination
around the MXC3 locus, a major gene for resistance to poplar leaf rust. Theor Appl Genet
103:1129–1137
Strauss SH, Lande R, Namkoong G (1992) Limitations of molecular-marker-aided selection in
forest tree breeding. Can J For Res 22:1050–1061
Telfer EJ, Stovold GT, Li YJ, Silva OB, Grattapaglia DG, Dungey HS (2015) Parentage recon-
struction in Eucalyptus nitens using SNPs and microsatellite markers: a comparative analysis
of marker data power and robustness. PLoS One 10:e0130601
Thumma BR, Southerton SG, Bell JC, Owen JV, Henery ML, Moran GF (2010) Quantitative trait
locus (QTL) analysis of wood quality traits in Eucalyptus nitens. Tree Genet Genomes
6:305–317
Van Eenennaam AL, Weigel KA, Young AE, Cleveland MA, Dekkers JCM (2014) Applied
animal genomics: results from the field. Annu Rev Anim Biosci 2:105–139
VanRaden PM (2008) Efficient methods to compute genomic predictions. J Dairy Sci
91:4414–4423
White TL, Adams WT, Neale DB (2007) Forest genetics. CABI Publishing, Cambridge, MA, p
682
Wilcox PL, Amerson HV, Kuhlman EG, Liu BH, O’Malley DM, Sederoff RR (1996) Detection of
a major gene for resistance to fusiform rust disease in loblolly pine by genomic mapping. Proc
Natl Acad Sci U S A 93:3859–3864
Williams CG (1988) Accelerated short-term genetic testing for loblolly-pine families. Can J For
Res 18:1085–1089
Williams CG, Neale DB (1992) Conifer wood quality and marker-aided selection: a case-study.
Can J For Res 22:1009–1017
Wright S (1931) Evolution in Mendelian populations. Genetics 16:97–159
Zapata-Valenzuela J, Isik F, Maltecca C, Wegrzyn J, Neale D, McKeand S, Whetten R (2012) SNP
markers trace familial linkages in a cloned population of Pinus taeda – prospects for genomic
selection. Tree Genet Genomes 6:1307–1318
Zapata-Valenzuela J, Whetten RW, Neale D, McKeand S, Isik F (2013) Genomic estimated
breeding values using genomic relationship matrices in a cloned population of loblolly pine.
G3 Genes Genom Genet 3:909–916
Zeng J, Toosi A, Fernando RL, Dekkers JCM, Garrick DJ (2013) Genomic selection of purebred
animals for crossbred performance in the presence of dominant gene action. Genet Sel Evol
45:11
Zhou LC, Holliday JA (2012) Targeted enrichment of the black cottonwood (Populus trichocarpa)
gene space using sequence capture. BMC Genomics 13:703
Index
A full conditional distributions, 69, 70

Alfalfa (Medicago sativa), 138 genome prediction, 86
Asexual propagation, 185 GBLUP, 58
Autohexaploid sweet potato, 186 GEBV accuracy, 57
Gibbs sampler (see Gibbs sampler)
GLMMs, 56
B heritability and training population size, 57
Bacterial artificial chromosome (BAC), 191 MCMC algorithm, 59
Barley, 100, 101, 109, 112, 119, 125 methodological developments, 86
Bayesian estimation, 39–52 non-Gaussian phenotypes, 55
Bayes’ theorem, 38, 39 observed and ordinal categorical
determination, parameters, 38 phenotypes, 56
Gibbs sampling (see Gibbs sampling) outcomes
linear mixed models (see Linear models) FHB data set, 83, 84
parameters, normal distributions GLS data set, 79–82
condition, 39 real data sets, 78–79
conjugate priors, 43 simulated data sets, 78
mean with known variance, 40, 41 parameters, 55
plausible values, 40 SNP marker panels, 55
probability density functions, 39 statistical, 61
variance with known mean, 41–43 applications, 59
Bayesian genomic-enabled prediction models, BLOR (see Bayesian logistic ordinal
59–62, 65–67, 70, 71, 78–84, 86 regression (BLOR) model)
accuracy, 56 BNBR, 62
BayesB method, 57 Gibbs sampler, 67
BayesC threshold model, 57 linear predictor, 61
binary trait, 58 Pois, 62
BLOR model (see Bayesian logistic ordinal polygenic effects, 66
regression (BLOR) model) regression coefficients, 65
BNBR (see Bayesian negative binomial threshold effects (γc), 66
regression (BNBR)) variance of polygenic effects, 66
data sets TGBLUP, 58
Fusarium head blight data, 61 threshold methods, 58
GLS and Septoria, 60 transformation, 56
EBV accuracy, 57 Bayesian LASSO methods, 194

DOI 10.1007/978-3-319-63170-7
252 Index
Bayesian logistic ordinal regression (BLOR) Pólya-Gamma distribution, 68

model, 78–80, 88–92 probability distribution, 68
CGGLUP, 59 scalar parameters, 81
cumulative logit model, 64 standard deviations (SD), 72
data augmentation approach, 59 Bayessian LASSO, 188
distributions, 61, 64 Best linear unbiased prediction
exclusive and exhaustive categories, 63 (BLUP), 164, 188
fit the real data set, 79 Breeding cycle, 187, 192, 194
full conditional distribution Breeding strategies, clonal crops, 186–187
liabilities, 88, 89 Breeding strategy, cereals development, 117
polygenic effects (bh), 91 genomics and phenomics, increased
regression coefficients (β), 90 precision, 116, 117
variance, location effects, 92 germplasm development, introgression
variance, polygenic effects, 91 of alleles, 115, 116
GBLUP, 59 implementation, GS, 113
Gibbs sampler, 67 MAS, 113
GLS data set, 79 program
interactions, 60 breeders, 117
liabilities, 63, 65, 71 design, 17
linear predictor, 61 data and training populations, 117
mean and standard deviation values, 73 introgress exotic alleles, 117
order statistics, 64 selection, alleles, 117
ordinal data analysis, 59 timing of GS application, 113–115, 117
outcomes
simulated data sets, 78
GLS data set, 79–80 C
parameter vector, 64 Chickpea (Cicer arietinum), 139, 140
parameters, 71 Clonal crops
Pólya-Gamma distribution, 87 asexual propagation, 185
restriction, 87 breeding programs, 194–195
thresholds, 63 breeding strategies, 186–188
validation samples, 80 fruit and forest trees, 185
Bayesian methods, 194 generations, 192–193
Bayesian negative binomial regression genetic analysis, 186
(BNBR), 75, 76, 78, 80–84 genetic architecture, 191–192
assessing prediction accuracy, 76, 77 genetic effects, 189–190
average values (mean) and standard genetic studies, polyploid crops, 186
deviation (SD), 74 GS, 187–189
FHB data set, 85 heritability, 189–190
fit the real data set, 79 high natural heterozygosity, 185
full conditional distribution, 92–94 LD, 190
GLS data set, 80, 83 polyploid, 185
hyper-parameters, 68 potato breeders, 186
implementation, 59, 72–75 proof-of-concept, 193
interaction, 60 single mother plant, 185
linear predictor, 68 Combining abilities, hybrid breeding, 152–153
measures, predictive performance Conditions, 87
proper scoring rules, 75, 76 Crop development, 2–6
scoring rules, 75 agriculture land, 1
outcomes challenges, 2
FHB data set, 83, 84 complex traits, 6
GLS data set, 80–82 conventional breeding, 2
simulated data sets, 78 food demand, 1
Index 253
food prices, 1 E
food production systems, 1 Eucalyptus, 212
GS breeding, 229
breeding, 4, 5 genotype, 218
clonal crops and tree species, 5–6 GS, 219, 222
GEBVs, 2 hybrid breeding, 227
methodologies and models, 3, 4 populations, 215
QTLs/genes, 2 RAD sequencing, 218
quantitative traits, 2 size, 214
training population, 2 SNP arrays, 217, 218
mapping populations, 2 tropical, 237
NGS technologies, 2
poverty eradication, 1
QTL, 2 F
Cross-validated prediction accuracy of GS, Forest tree breeding, 201–203
hybrid breeding applications, 238
biparental populations, 156–158 breeding cycle, 238
concepts, 154 eucalyptus breeding, 238
estimation genomic selection (see Genome-wide
marker effects, 158 selection, forest tree breeding)
and validation, 155 standard phenotypic selection, 237
factorial crosses, 158–160 top ranked seedlings, GEBV, 238
limitations, 157 training dataset, 238
plant material, 155 trait dissection to genomic prediction,
predictabilities, 157 199–201 (see also Genomic
program, 155–158 prediction in forest trees)
same genotypes, 155 Fresh root weight (FRW), 193
segregating inbred populations, 155
test and target environments, 160–161
Cross validation, legumes, 134, 136, 138 G
Genetic architecture
traits and size of training population,
D 191–192
Disease-free stocks, 187 Genetic diversity
Dominance, hybrid breeding, 150–151 inbreeding and maintenance, 235, 236
Dominance within GS model Genetic effects
accommodating epistasis, 167–168 clonal crops, 189–190
Bayesian approaches, 165–167 Genetic studies in polyploid crops, 186
benefits, 166 Genome-based selection, 187
genotype environment interactions, 170 Genome-wide selection, forest tree breeding,
genotypes and relationship matrices, 203, 204, 206, 215–221, 228–230,
164–164 233–235
marker effects changes, environment and model
BayesA, 162–163 retraining, 241
BayesB, 163 complex traits, 203
BayesCπ, 163–164 cost-benefit analysis, 240, 241
components, 161 cultivated plant species, 203
linear unbiased prediction, 162 data analysis, 240
metabolomic prediction, 169–170 marker variables, 228
multiple-trait, 168–169 modeling nonadditive effects, 229
W-BLUP, 161 prediction of genome-estimated
Dry matter content (DMC), 193 breeding, 228
254 Index
Genome-wide selection (cont.) prediction, generations, 241

ranking problem, 229, 230 stages, development and
risk, 228 implementation, 205
RR-BLUP method, 228 training GS models, 239
dense marker data and QTL effects, 202 trait heritability and genetic architecture,
EBV, 201 227, 228
effective population size, 213–215 training population, 202
environments and ages, 230–233, 239 validation\population, 202
experimental outcomes, 206–212 Genome-wide SNPs, 191
features and outcomes, 207–211 Genomic best linear unbiased prediction
G-BLUP, 202 (G-BLUP), 14, 15, 188, 194
GEBV, 202 Genomic estimated breeding values (GEBVs),
genetic relationship, training population 2, 187, 190, 192
and selection candidates, 239 Genomic prediction, legumes, 134
genetic values, 201 Genomic prediction in forest trees, 230–233
genomic prediction, 203 alleles, traits, 199
genotyping, 239 challenges, tree breeders, 199
genotyping density and SNP dissecting complex traits, 200
DNA marker, 215 environments and ages
fixed content, 217, 218 accuracy, 231, 232
GbS approaches, 218–221 crops, 233
GEBV accuracy, 216 GEBV, 231
GEBV prediction accuracy, 216 interaction, 231
LD pattern, 217 management zones, 231
lower-density marker panels, 217 marker effect, 233
microsatellite markers, 216 outcomes, 232
molecular genetic techniques, 215 periodical retraining, 233
population size, 216 ranking of individual trees, 230
whole-genome sequence data, 221 same clones or families, 231
inbreeding and maintenance of genetic selection, same age, 232
diversity, 235, 236 tree breeding programs, 230
logistic issues, 240 variation, 231
marker effect, 202 genomic selection, 201
MAS, 202 growth and wood quality, 200
performance MAS, 200
breeding program, 235 modeling nonadditive effects, 229
G0, G1, and G2, 234 QTL and association mapping, 200
long-term gain, 233 ranking problem, 229, 230
Pinus pinaster, 234 recurrent cycles of selection, mating and
training and validation sets, 233 testing, 199
perspectives tree breeding programs, 199
accuracy, 204 Genomic prediction models
BLUP, 206 across environment vs. within environment
EBV and true breeding value, 203 replication, 18, 19
eucalypts and poplars, 204 breeding programs, 16
GEBV/GEGV, 204 MAS, 16
genetic improvement, 203 preliminary and advanced breeding tests,
implementation, 204 17, 18
measures, tree life, 203 QTL mapping, 16
phenotypes, 203 replication and plot allocation, 17
training population, 204 replication, phenotypic variance, 16
tree breeding program, 203, 204 Genomic selection (GS), 2, 185
Index 255
candidate population, 2 objective function, 171–173

clonally propagated crops (see Clonal parameters, 175–176
crops) plant breeding programs, 171, 173, 177
crop (see Crop development) predictability, 171, 177
GEBVs, 2, 3, 187 recalibration, 176–178
genome-wide marker data, 2 International Institute of Tropical Agriculture
methodologies and models, 3, 4 (IITA), 191
Genomics-assisted breeding (GAB), 132, International Potato Center (CIP), 191
133, 136
Genotype-by-environment (G x E)
interactions, 186, 192 L
Gibbs sampler Legume crops
BNBR Alfalfa, 138
parameters, 70 breeding, 132
simulation, 71 chickpea, 139, 140
BLOR, 67 cultivation, 131
Gibbs sampling demand, 132
Box-Muller method, 50 GAB approaches, 133
conditional probability distributions, 49 GBS approach, 133
parameters estimation, 50–52 generation, genomic resources, 141
procedure, 49 genomic-assisted breeding, 141
values, 49 genomic prediction, 134
Gray leaf spot (GLS), 60 genomic resources, 132
Groundnut (Arachis hypogaea), 140–141 genomic selection studies, 137
groundnut, 140
GS, 134–136, 141
H introgression lines, 133
Hardy–Weinberg equilibrium, 190, 192 large scale genome-wide genetic
Heritability, clonal crops, 189–190 markers, 132
Heterosis, hybrid breeding, 149, 150, 152–154 MABC, 133, 134
High-density genome-wide molecular MARS, 133, 134
markers, 187 NGS technologies, 132
Hybrid breeding pea, 139
combining abilities, 152–153 production, 131
concept of dominance, 151–152 productivity trends, 132
GS QTLs, 134
challenges, 178 single nucleotide polymorphisms
cross-validation (see Cross-validated (SNPs), 132
prediction accuracy, GS) soybean (Glycine max), 136–138
dominance (see Dominance within GS UN World Food Program, 131
model) Linear models, 44, 46–49
implementation (see Implementation Bayesian estimation
of GS in hybrid breeding) conditional probability distribution β,
skills and knowledge, plant 46, 47
breeders, 178 conditional probability distribution, 49
heterosis, 149–150 conditional probability distribution, uj,
heterotic group and patterns, 153–154 47, 48
line breeding, 151 parameters, 44
prior probability distributions, 44
quantitative traits, 25–27
I Linkage disequilibrium (LD), 189
Implementation of GS in hybrid breeding markers and quantitative trait loci, 190
biparental population, 173, 174, 176, 178 QTL mapping, 10
budget, 171, 173, 174 Linear models, 43, 45, 46
decision variables, 174–175 Bayesian estimation
256 Index
Linear models (cont.) deviations, 31

joint posterior probability distributions, genetic factors, 31
45–46 genetic markers, 32, 33
MCMC methods, 43 GS, 32
prior probability distributions, 45 genotypic value, 32
LD, 33
narrow-sense, 32
M traditional genomic selection, 32
Marker-assisted backcrossing (MABC), 133 values, 31
Marker-assisted selection (MAS), 187 decomposition of variance and genetic
crop development, 4 markers, 30, 31
oats, 109 genetic factors, 23
wheat, 101 genetic markers, 52
MARS. See Marker-assisted recurrent selection GS
(MARS) dataset, 24
genetic markers, 24
process, 23
N strategies, 23
Next-generation sequencing (NGS) LD, 53
technologies, 2 least-squares estimation
genetic markers, 34
number of parameters, 37, 38
O one-marker model, 34, 35
Oats, 100, 109 parameters, 33
two-marker model without
interactions, 35–37
P linear models, 25–27
Pea (Pisum sativum), 139 one-marker model, 27, 28
Phenotypic data, 87 prediction, 52
Phenotypic selection (PS), 188, 193 two-marker model
Picea, 214, 217, 218 with interactions, 29
Pinus without interactions, 28–29
generations, 212
genotyping, 218
SNP arrays, 217 R
Population size, legumes, 135, 136, 138–140 Random regression-best linear unbiased
Potato breeders, 186 prediction model (RR-BLUP), 193,
Prediction accuracy, 5, 6, 188, 189 194
Public sector, 193 Reproducing kernel Hilbert space (RKHS),
188, 194
Ridge-regression best linear unbiased
Q prediction (RR-BLUP), 14, 17, 188
Qualitative and quantitative trait loci Rye, 100, 101
(QTL), 187
Quantitative Trait Loci
LD, 190–191 S
Quantitative traits, 23, 24, 28, 29, 31–52 Shoot weight (SW), 193
Bayesian estimation (see Bayesian Single mother plant, 185
estimation) Small grains, 101, 109–125
breeding values and heritability breeding methods (see Breeding strategy,
additive genotypic value, 32 cereals development)
broad-sense, 32 breeding programs, 125
Index 257
GS diverse germplasm collection, 12–14

cost benefit analysis, 124, 125 family-specific, 9
pure line breeding program, 114 G-BLUP perspective, 14
studies, 102–105, 107 genetic relationships, 15
GS prediction genomic prediction, 9, 10, 14
accuracies, 110 genomic selection, 8, 9
effect, number of markers and genotyping, 10
individuals, 112 LD source of information, 15
model, 111 maize breeding, 9
research, 110 marker-based optimization, 10
training and validation population, 111, marker-QTL LD, 15
112 optimization, 9
GS research phenotypic data, 8
DArT, 109 QTL mapping, 10
GBS, 109 researchers, 14
oat breeding program, 109 RR-BLUP perspective, 14
oats, 109 sizes and marker numbers, 10
wheat, 101 unrelated genotypes, 11, 12
GxE development, 8
applications, GS models, 118–119 genetic value, 7
environmental covariates and crop genotyping costs, 8
models, 123, 124 hypothetical and empirical
modeling marker replication and implementation, GS, 8
interaction, 119–121 literature, 19
phenotypic variation, 118 molecular markers, 7
TPE, 118 optimization, 121
treating environments, multiple traits, phenotypic testing procedures, 7
121, 122 phenotyping, 7, 19
marker effects, 120 resource allocation, phenotyping, 16–19
optimization, training population, 121 tree breeding
Small grains clonal replication, 223
biparental cross of maize, 100 eucalyptus, 222
breeding cycle, 100 GS in hybrid breeding, 226, 227
breeding program, 100 NIRS, 222
cereals, 100 phenotyping, 222
estimated breeding values, 100 QTL distribution, 222
genotype-by-environment interaction selection accuracy, 222
(GxE), 101 and selection candidates, 223–226
pseudo-cereals, 100 SNP effects, 221
RR-BLUP, 101 Trait heritability and genetic
Soybean (Glycine max), 136–138 architecture, 227, 228
Starch yield (SY), 193 Triticale, 100, 116
T V
Target population of environments (TPE), 118 Vegetative propagules, 187
Time consuming and resource consuming, 186
Training population, 8–15
breeding programs, 7 W
design Wheat
composition, 8, 9 bread, 101
co-segregation of alleles, 15 breeding programs, 101, 112
258 Index
Wheat (cont.) lines, 123

buckwheat, 100 low-intensity winter, 124
durum wheat, 101, 120 MET, 122
environmental covariates, 122 phenotypes, 120
GBS datasets, 109 quality, 125
GS, 100, 109 SirusQuality, 123
investigation, 111

Genomic Selection For Crop Improvement - New Molecular Breeding Strategies For Crop Improvement (Springer, 2017)

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Genomic Selection For Crop Improvement - New Molecular Breeding Strategies For Crop Improvement (Springer, 2017)

Uploaded by

Copyright:

Available Formats

Rajeev K.

Genomic Selection for Crop

ISBN 978-3-319-63168-4 ISBN 978-3-319-63170-7 (eBook)

Library of Congress Control Number: 2017951639

© Springer International Publishing AG 2017

Printed on acid-free paper

This Springer imprint is published by Springer Nature

Today the world is facing an unprecedented challenge: how to feed a growing

ICRISAT Dr. David Bergvinson

Patancheru, Telangana, India Rajeev K. Varshney

1 Genomic Selection for Crop Improvement: An Introduction . . . . . 1

9 Status and Perspectives of Genomic Selection in Forest

Jared Crain Department of Plant Pathology, Kansas State University, Manhattan,

Manish K. Pandey International Crops Research Institute for the Semi-Arid

Rajeev K. Varshney, Manish Roorkiwal, and Mark E. Sorrells

R.K. Varshney (*) • M. Roorkiwal

© Springer International Publishing AG 2017 1

1.2 Methodologies and Models for GS

1.3 GS in Field Crop Breeding

for other legumes to start deployment of GS in those breeding programs to achieve a

1.4 GS for Improvement of Clonal Crops and Tree Species

reduced by genotyping young seedlings and predicting their phenotype instead of

Aaron Lorenz and Liana Nice

Obtaining accurate and inexpensive estimates of genetic value is a fundamental

A. Lorenz (*) • L. Nice

© Springer International Publishing AG 2017 7

with measured phenotypes to make selections in the breeding program (Johnson

2.2 Training Population Design

2.2.1 Training and Target Populations Are Segregating

The most straightforward way to conduct genomic selection is to create family-

2.2.2 Training and Target Populations Include Related

2.2.3 Training and Target Populations Include Lines from

Besides predicting the genetic value of progenies comprising an active breeding

large, containing up to hundreds of thousands of plant accessions. Advancements in

2.2.4 Sources of Information and Population Genomic

could be more uniformly distributed among individuals in training populations that

2.3 Resource Allocation for Phenotyping for Genomic

A key design aspect of breeding programs is the allocation of resources among

sufficiently addressed by previous MAS studies. Specifically, we review: (1) the

2.3.1 Replication and Plot Allocation for Calibrating

To determine whether resource allocation recommendations for MAS can be

2.3.2 Allocation of Resources Across Preliminary

2.3.3 Across Environment Versus Within Environment

Schopp P, Muller D, Technow F, Melchinger A (2017) Accuracy of genomic prediction in

Akihiro Nakaya and Sachiko Isobe

© Springer International Publishing AG 2017 23

3.2 A Dataset for Genomic Selection

Fig. 3.1 A dataset example for genomic selection

3.3 Prediction Models

3.3.1 Linear Models

where R is a known matrix of N N dimensions. When we assume that the total

3.3.2 One-Marker Model

We first consider a simple example where no environmental factors (L ¼ 0) and

Here, yi ¼ β0 + u11 + ei if its genotype is γ 1, and yi ¼ β0 + u12 + ei if γ 2. Letting

3.3.3 Two-Marker Model Without Interactions

yi ¼ β0 þ u11 þ u21 þ ei , ð3:9Þ

Letting a0 ¼ β0 + u11 + u21, a1 ¼ u12  u11, and a2 ¼ u22  u21, we have

yi ¼ a0 þ a1 g1i þ a2 g2i þ ei , ð3:13Þ

3.3.4 Two-Marker Model with Interactions

Here, the columns of W correspond to the combinations of the genotypes of the

3.4 Decomposition of Variance and Contributions

The linear model given by Eq. 3.1 is rewritten as follows:

V ðyÞ ¼ V ðXβÞ þ V ðZ1 u1 Þ þ V ðZ2 u2 Þ þ þ V ðZM uM Þ þ V ðeÞ: ð3:18Þ

Letting a0 ¼ β0 + u11 + u21, a1 ¼ u12 u11, and a2 ¼ u22 u21, we have

Equation 3.31 can be written as ∂J/∂a1 ¼ 2N(σ y1 a1σ 11) ¼ 0. Here, σ y1 is

J ¼ ðGa yÞ0 ðGa yÞ: ð3:45Þ

J ¼ ða0 G0 y0 ÞðGa yÞ ¼ a0 G0 Ga 2a0 G0 y þ y0 y: ð3:46Þ

∂J=∂a ¼ 2G0 Ga 2G0 y: ð3:47Þ

a ¼ ðG0 GÞ1 G0 y: ð3:49Þ

πðuj jβ, uj , v, σ 2e , yÞ

πðuj jβ, uj , v, σ 2e , yÞ

uj j β, uj , v, σ 2e , y Nðuj , Σj Þ: ð3:90Þ

σ 2e j β, u, v, y Scaledinvχ 2 ðN, s2e Þ, ð3:94Þ